If you're learning how to create a prompt that consistently produces the best results, this guide is for you. Prompt engineering is the skill of communicating effectively with AI models. Whether you're using ChatGPT, Claude, Gemini, or any other LLM, how you phrase your request dramatically affects the quality of the output.

Why Prompt Engineering Matters

The same AI model can give you a mediocre paragraph or a brilliant, structured response — the difference is the prompt. Good prompts lead to:

• Better accuracy — the AI understands exactly what you want
• Consistent quality — reliable outputs every time
• Time savings — fewer back-and-forth iterations
• Professional results — content you can actually use

The 5-Part Prompt Structure

Every great prompt follows a basic structure. Think of it as a recipe:

1. Role

Tell the AI who it should be. This sets the expertise level and perspective.

You are a senior content strategist with 10 years of experience in B2B SaaS marketing.

2. Context

Provide background information the AI needs to understand your situation.

I'm launching a new AI writing tool targeted at freelance writers. Our main differentiator is the domain-specific templates.

3. Task

Be specific about what you want the AI to do. Use action verbs.

Write 5 LinkedIn post ideas that highlight our template feature. Each post should be 150-200 words, conversational, and end with a CTA.

4. Constraints

Set boundaries: tone, length, format, what to avoid.

Tone: friendly but professional. Avoid jargon. Don't use more than 2 emojis per post. No hashtag overload.

5. Format

Specify the output structure you need.

Return as a numbered list. Each item should have: a hook line, the body, and the CTA in bold.

Advanced Techniques

Chain-of-Thought (CoT)

Ask the AI to think step-by-step before giving the answer. This dramatically improves reasoning tasks.

Think step-by-step before answering. First analyze the data, then identify patterns, then give your recommendation.

Few-Shot Learning

Give the AI 2-3 examples of what you want. This is the fastest way to teach it your style.

Here are 2 examples of the output I want:

Example 1: [your example]
Example 2: [your example]

Now generate a new one following the same pattern.

Output Iteration

Don't try to get the perfect result in one shot. Use follow-up prompts to refine:

Good, but make it more conversational. Shorten the paragraphs. Add a specific example for the second point.

Common Mistakes to Avoid

• Being too vague: "Write me something about marketing" vs "Write a LinkedIn post about why short-form video outperforms static images for B2B brands"
• Too many instructions: Break complex requests into smaller steps
• No examples: Show, don't just tell. Examples are worth 1000 words of instruction
• Ignoring output format: Always specify how you want the result structured

Putting It All Together

Here's a complete prompt using the 5-part structure:

# Role
You are a copywriter specializing in SaaS landing pages.

# Context
I'm building a landing page for PromptLab, a free AI prompt workspace. The target audience is marketers and content creators who struggle with writing effective prompts for ChatGPT and Claude.

# Task
Write the hero section copy for the landing page, including:
- A headline (max 8 words)
- A subheadline (max 25 words)
- A primary CTA button text (3-5 words)

# Constraints
- Tone: confident, clear, slightly playful
- No buzzwords like "revolutionary" or "game-changing"
- Focus on the pain point (bad prompts = bad AI output)

# Format
Return as:
**Headline:** ...
**Subheadline:** ...
**CTA:** ...

Or — even easier — just paste this into PromptLab's Builder, pick "Marketing" as the category, and let the engine generate a structured prompt with all 5 components automatically.

ChatGPT and Claude are the two most popular AI assistants — but they respond very differently to the same prompt. Understanding these differences is the key to getting better results from both.

The Core Difference

ChatGPT (GPT-4o) tends to be more creative, verbose, and eager to please. It follows instructions literally and produces longer outputs by default.

Claude (Sonnet/Opus) is more careful, nuanced, and prefers structured instructions. It excels at following complex multi-step instructions and maintaining context.

Prompting for ChatGPT

ChatGPT works best with:

• Direct, action-oriented instructions: "Write 5 variations of..."
• Role-playing: "Act as a senior developer reviewing code"
• Delimited input: Use triple backticks for code or data
• Explicit format requests: "Return as a markdown table"

You are a social media expert. Write 5 Instagram captions for a fitness app launch. Each caption should include:
- A hook in the first line
- 3-4 value points
- A CTA with emoji
- 5 relevant hashtags
Tone: energetic, motivational, Gen Z friendly.

Prompting for Claude

Claude responds better to:

• Structured sections: Use headers (# Role, # Context, etc.)
• XML tags: Claude is trained to respect XML boundaries
• Detailed constraints: Claude excels at following complex rules
• Examples: Few-shot works exceptionally well with Claude

You are a content strategist. Analyze the following blog post and suggest 3 improvement areas.



[Paste your blog post here]



For each improvement:
- What needs to change and why
- A specific rewrite suggestion
- Impact level: High / Medium / Low

Quick Reference

• Creative tasks: ChatGPT tends to be more inventive
• Analytical tasks: Claude tends to be more thorough
• Long conversations: Claude maintains context better
• Code generation: Both are excellent; Claude is more careful
• Structured output: Claude follows format instructions more precisely

The best way to learn is to experiment. Open PromptLab, create a prompt, and compare the results. The readiness score will tell you if your prompt works well for both models — or if it needs adjustment.

Knowing how to create the best prompt for marketing is the difference between AI copy that sounds robotic and copy that actually converts. Here are 10 battle-tested templates you can copy, paste, and customize in PromptLab.

1. Social Media Caption

Write 5 Instagram captions for [product/announcement].
- Hook line (attention-grabbing)
- 3 value points
- CTA with emoji
- 5 hashtags
Tone: [brand voice]

2. Email Subject Lines

Generate 10 email subject lines for [campaign purpose].
Include: 3 curiosity-based, 3 urgency-based, 2 benefit-driven, 2 personalized.
Keep under 50 characters each.

3. Blog Outline

Create a detailed blog outline for: [topic]
Target audience: [audience]
Goal: [awareness/consideration/conversion]
Include: title, introduction hook, 5-7 H2 sections with bullet points, conclusion with CTA.
Word count target: [X] words.

4. Ad Copy (Facebook/Google)

Write 3 variations of ad copy for [product].
Headline (max 40 chars), primary text (max 125 chars), description (max 30 chars).
Each variation targets: 1) pain point, 2) benefit, 3) social proof.
CTA: [shop now / learn more / sign up]

5. Landing Page Hero

Write a landing page hero section for [product].
- Headline: max 8 words, punchy
- Subheadline: max 25 words, explain the value
- Primary CTA: 3-5 words
- Secondary CTA: "See How It Works"
Tone: [professional / playful / authoritative]

6. Newsletter Introduction

Write a newsletter intro for this week's edition.
Main topic: [topic]
Key takeaway: [one sentence]
Tone: conversational, like writing to a friend.
Include a transition to the first article.
Max 100 words.

7. Product Description

Write a product description for [product name].
Features: [list features]
Target buyer: [persona]
Include: headline, 3 benefit bullets, emotional hook, CTA.
Optimize for: clarity, scannability, conversion.

8. A/B Test Hypotheses

Generate 5 A/B test hypotheses for [page/campaign].
For each, specify: variable, hypothesis (if we X, then Y because Z), expected impact, priority (high/med/low).

9. Competitor Analysis

Analyze [competitor name] vs our [product].
Compare: positioning, messaging, pricing, target audience, strengths, weaknesses.
Output: comparison table + 3 opportunities we can exploit.

10. Content Repurposing

Repurpose this [blog post/transcript/video] into:
1. 5 tweet threads
2. 1 LinkedIn post (200 words)
3. 3 Instagram carousel slides
4. 1 email newsletter (300 words)
Original content: [paste here]

If you want to know how to create a prompt that actually delivers — one that gets you a usable draft on the first pass instead of a vague answer you'll rewrite — start here. These 9 rules are the shortest path to the best prompt for Claude, distilled from Anthropic's own prompt engineering guide and battle-tested across thousands of real conversations.

Writing prompts is engineering, not magic. The difference between a vague "help me with this" and a sharp, testable instruction is the difference between a draft you'll throw away and one you can ship. These nine rules are the shortest path to the second kind.

The 9 Rules

1. Name the Output, Not the Task

Replace weak verbs like "review", "help", "look at", "improve" with a specific deliverable: a table, a JSON object, a five-bullet list, a doc, a 200-word summary, three Slack-ready messages.

Why: Vague verbs produce vague drafts. The model has to guess what "good" looks like. If you can't name the output, the model can't either.

Bad: "Help me with this landing page."

Good: "Audit the landing page above. Return a markdown table with three columns: Element, Issue, Fix. Cover hero, CTA, social proof, and footer. No preamble, no recap."

2. Define the Length Up Front

State the count, the word budget, or the structural shape before the model starts writing. "5 bullets" beats "a few bullets." "180 words" beats "a paragraph." "Three sections, first section is the hook" beats "an intro."

For lists, name the first word of each line so the model can parallelize. For prose, add: "No preamble. No recap. No filler."

Why: Without length, models default to verbosity. With length, they constrain themselves. "Five paragraphs" is a different prompt from "write this."

3. Flip Every "Don't" Into a "Do"

Find every don't, avoid, never, without in your prompt. Rewrite each as a positive instruction. Models follow what they should do far more reliably than what they shouldn't.

Bad: "Don't use jargon, don't be vague, don't be preachy."

Good: "Use plain language. Be specific with numbers. State the benefit in one line, then back it with evidence."

Why: Claude 4.7 reads instructions literally. A "don't" tells the model what to filter out; a "do" tells it what to generate. Always bias toward generation.

4. Lead With Action

Strip the throat-clearing. "Can you help me with..." "I'd like you to..." "I need..." — all of these waste the first ~30 tokens of context.

Start with a verb: Write, Draft, Audit, Convert, Generate, List, Summarize, Rewrite, Translate, Score.

Bad: "I was wondering if you could maybe help me think about how to structure a Q3 OKR doc?"

Good: "Draft a Q3 OKR doc. Three objectives, each with 3 key results. Use the SMART format. Audience: CEO + leadership team. 400 words."

Why: The model's first tokens are the most expensive (cache + attention). Spend them on the work, not on politeness.

5. Force Maximum Reasoning

For non-trivial tasks, select the strongest reasoning model and explicitly ask for it. In Claude 4.7, that means Opus with Adaptive Thinking turned on.

Add: "Think before answering. State the assumptions. Walk through the reasoning. Then give the final answer."

For simple, well-defined tasks, do the opposite — turn reasoning off, because you want speed, not analysis paralysis. Claude 4.7's reasoning toggle is your friend.

Why: Reasoning effort is a parameter, not a vibe. You can over-reason a one-line answer and under-reason a strategic decision. Match the tool to the task.

6. Add "Go Beyond the Basics"

For creative and strategic work, ban the lazy defaults. Tell the model: "Don't give me the obvious answer. Pretend I'm a real client who has seen the generic version already. Go one layer deeper."

This single line changes output quality more than any other trick. Pair it with: "List 3 contrarian takes. Then pick the strongest and defend it."

Why: LLMs are trained to be helpful, which defaults to safe, which defaults to generic. The "go beyond" instruction breaks the gradient and unlocks the tail of the distribution.

7. Upload Your Voice

Paste 2-3 sentences of exactly how you (or your brand) sounds. Then add: "Match the style of these examples. Don't tighten it. Don't formalize it. Keep the same rhythm."

Save this as a reusable "about-me" file in PromptLab — paste it once, reference it forever. Voice is the hardest thing for models to nail from instructions alone. Examples are 10x cheaper than adjectives.

Why: "Professional but warm" is meaningless. "Short sentences. Em-dashes. No exclamation marks. Starts with the punchline" is a prompt.

8. Control Tools On Purpose

Decide upfront whether you want the model to use tools — and which ones.

• Research tasks: "Search the web aggressively. Cite at least 2 sources per non-trivial claim. Quote exact numbers."
• Speed tasks: "Answer from training. No search. No tools. Go."
• Workflow tasks: Name the connectors explicitly: "Use Gmail to find the last 3 threads with Acme Corp. Use Calendar to check availability next week. Draft a reply in Slack-ready format."

Why: Claude 4.7 (and GPT-5) call fewer tools by default than 3.5 did. If you want web search or a connector fired, you have to ask. If you don't, say so — otherwise the model burns time and tokens deciding.

9. State the Goal Before the Task

Open the prompt with the win condition, not the workflow.

Bad: "Write me a follow-up email."

Good: "Goal: Get a meeting booked with the Head of Growth at Acme Corp by Friday. Audience: VP of Marketing, 15 years experience, skeptical of cold outreach. Output: 3 follow-up email variations under 80 words each. Subject line under 45 chars."

Name the audience (CRO, not engineer), the deadline, the measurable outcome. A prompt without a goal is a wish. A prompt with a goal is a brief.

Why: The model can trade off tone, length, and depth intelligently — but only if it knows what success looks like. Without a goal, it optimizes for the average of the training data. With a goal, it optimizes for your outcome.

Putting It All Together

Here's the template. Save it in PromptLab as a starter:

Goal: [what winning looks like in one sentence]
Audience: [who reads this, their seniority, their skepticism]
Output: [format — table, list, doc, JSON, code]
Length: [count, word budget, or section structure]
Voice: [paste 2-3 sentences of exactly the tone you want]
Rules: [the "do" version of every "don't"]
Tools: [search? connectors? or none?]
Reasoning: [on or off, and why]
Go beyond the obvious: [the "go deeper" instruction]
Now: [the actual task, starting with a verb]

That's it. Nine rules, one template, no magic words. The difference between a prompt that gets ignored and a prompt that ships is almost always structural — and structure is a skill, not a talent.

Want to see the template in action? Open PromptLab, paste the template into a new prompt, and ship your first structured prompt in under 3 minutes.

Most people type two sentences into ChatGPT, get a mediocre answer, and assume the model is "just not that smart." The model is not the bottleneck. The prompt is. The difference between a flat paragraph and a sharp, usable response is almost always structural — and once you learn the structure, the same free model that gave you a C-minus draft will hand you an A.

This guide is the practical, no-fluff version of everything OpenAI's own prompt engineering documentation teaches — condensed into one framework, 12 copy-paste prompts, and the six mistakes that quietly kill ChatGPT output quality. If you've ever wondered how to create a prompt for ChatGPT that actually delivers, this is for you.

Why Most ChatGPT Prompts Underperform

ChatGPT is a generalist. It was trained to be helpful, harmless, and broadly useful across billions of conversation patterns. That training pushes the model toward the statistical middle of any response: safe, balanced, vaguely confident, and rarely surprising.

A good prompt breaks that gradient. It tells the model three things at once: who it is (role), what situation it's writing for (context), and exactly what shape the answer should take (format and constraints). Skip any of those three and you get the average answer. Hit all three and you get the answer you would have written yourself on a good day — in 4 seconds.

The CRISPE Framework: The 6-Part Prompt Recipe

The most consistent framework we've tested across thousands of ChatGPT prompts is CRISPE — an acronym for Capacity, Request, Insight, Statement, Personality, Experiment. It works because it forces you to specify everything the model needs before it starts predicting tokens.

1. Capacity & Role

Tell ChatGPT who it is. A role sets the expertise level, vocabulary, and default frame for the response. Without one, the model defaults to "a helpful assistant," which is the lowest-energy version of itself.

You are a senior conversion copywriter who specializes in B2B SaaS landing pages. You write in the style of Harry Dry (MarketingExamples) — punchy, specific, no fluff.

2. Request (the actual task)

The verb-forward, unambiguous ask. Lead with the action. Skip the "could you please" — it costs tokens and adds nothing.

Write 5 hero-section headlines for a prompt-engineering tool called PromptLab. Each headline must be under 8 words, focus on the pain (writing bad prompts = bad AI output), and avoid the words "revolutionary," "powerful," and "ultimate."

3. Insight (background context)

Everything the model needs to know that isn't in the request itself: audience, product, stage, what was tried before, what's at stake.

Context: PromptLab is a free AI prompt workspace. Target audience is marketers and content creators who already use ChatGPT daily but get inconsistent results. Main differentiator: a Readiness Score that scores any prompt on structure. Competitor copy from Notion AI and Copy.ai is too generic — we want sharper, more opinionated.

4. Statement (the mission / success criteria)

What does "winning" look like? This is the line most prompts skip — and it's the single highest-leverage sentence you can add.

Goal: a headline that, when read alone, makes a marketer stop scrolling and click. Optimized for clarity and curiosity, not cleverness.

5. Personality (tone, voice, style)

Name the voice with examples, not adjectives. "Professional but warm" is meaningless. "Short sentences. Em-dashes. No exclamation marks. Starts with the punchline" is a prompt.

Voice: short sentences. Em-dashes allowed. No exclamation marks. Starts with the verb or the pain, never the brand. Think: Alex Hormozi meets Paul Graham.

6. Experiment (format + iteration instructions)

How should the answer be shaped? Markdown table? JSON? Numbered list? 200 words? And what should the model do if it doesn't know — should it guess, ask, or refuse?

Format: return as a numbered list. Each item is the headline on its own line, no commentary. If you would need more than 8 words, stop and ask before submitting.

12 Copy-Paste ChatGPT Prompts That Actually Work

Steal these. Customize the bracketed parts. Each one is built on CRISPE and tested across real ChatGPT conversations.

1. The "Think Step-by-Step" Reasoning Prompt

For analysis, planning, and decisions. Adding a "let's think step by step" instruction measurably improves accuracy on multi-step problems — this is one of the most replicated findings in prompt engineering research, originally published in the chain-of-thought paper from Google and Princeton.

You are a senior strategy consultant. I'm trying to decide whether to [X].

Before giving a recommendation, walk through the reasoning:
1. List the 3 strongest arguments FOR
2. List the 3 strongest arguments AGAINST
3. Identify the 2 unknown unknowns that could flip the decision
4. Give your final recommendation in one sentence, with confidence level (low/medium/high)

2. The Few-Shot Style Prompt

Show, don't tell. Give ChatGPT 2-3 examples of exactly the output you want, then ask for a new one. This is the fastest way to teach the model your style without writing a single adjective.

Match the tone and structure of these two example tweets:

Example 1: Stop prompting like a search engine. Start prompting like a brief.
Example 2: Your AI is only as smart as the question. Most questions are vague.

Now write 5 more tweets about [TOPIC] in the same voice. Each under 200 characters. No hashtags. No emojis.

3. The "Act As" Expert Prompt

The classic role prompt — use it when you need the model to draw on a specific domain's vocabulary and frame.

Act as a senior tax accountant with 15 years of experience advising freelancers in the US.

Answer the question below as if I were a new client sitting across from you. Use plain English. Flag any case where you'd need to look at my actual documents before answering.

Question: [YOUR QUESTION]

4. The Before/After Editor Prompt

When you already have a draft and want a sharper version, don't ask for "feedback" — ask for a rewrite with constraints.

Below is a draft I wrote. Rewrite it following these rules:
- Cut 30% of the words
- Lead with the conclusion, not the build-up
- Replace any adjective pair ("very unique," "really important") with one specific word
- No passive voice

Draft:
[PASTE YOUR TEXT]

5. The Structured Output Prompt

For data extraction, comparison, or anything you'd normally put in a spreadsheet.

Extract the following from the text below: company name, funding round, amount raised, lead investor, date announced.

Return as a markdown table with headers: | Company | Round | Amount | Lead Investor | Date |

If a field is missing, write "N/A" — never guess.

Text:
[PASTE TEXT]

6. The Persona Interview Prompt

For user research, customer development, and market positioning. This one mimics a real 1:1 conversation surprisingly well.

You are a [PERSONA — e.g., freelance designer, 4 years experience, $80k/year, uses ChatGPT weekly but feels it's hit-or-miss].

I'm going to ask you 5 questions about how you discovered, evaluated, and decided to use [PRODUCT/CATEGORY]. Answer in first person, with the specific pain points and tradeoffs a real user would name. Be honest — include at least one complaint per answer.

Questions:
1. [Q1]
2. [Q2]
3. [Q3]
4. [Q4]
5. [Q5]

7. The "Teach Me" Prompt

For learning any new concept faster than reading a textbook. Force the model to explain at three depths, then quiz you.

Teach me [TOPIC] at three levels:

Level 1 (ELI5): a 12-year-old should get it.
Level 2 (Practitioner): a smart colleague should get it.
Level 3 (Edge cases): an expert should learn something new.

Then give me 5 questions to test my understanding, mixed difficulty. Don't give me the answers — I'll respond and you grade me.

8. The Constraint-Removed Brainstorm Prompt

For ideation, when default brainstorming feels stale. Remove the "don't be crazy" filter and force the model to go wide before narrowing.

Generate 15 wildly different ideas for [PROBLEM]. Don't filter for feasibility yet.

Constraints:
- At least 5 must be ideas a Fortune 500 company could never do
- At least 5 must cost less than $100 to test this week
- At least 5 must be embarrassing enough that I'd hesitate to put them in a pitch deck

Then pick the 3 you think are most likely to actually work, and explain why in one line each.

9. The Code Review Prompt

For developers. Forces ChatGPT to review like a senior engineer would, not just describe what the code does.

Review the code below like a senior engineer doing a PR review. Specifically look for:
- Bugs or edge cases
- Performance issues
- Security vulnerabilities
- Readability / naming
- Anything that would make you leave a comment on the PR

Format each issue as: severity (blocker/major/minor), file location, one-line description, suggested fix.

If the code is good, say so. Don't pad.

Code:
[PASTE CODE]

10. The Email Reply Prompt

For inbox triage. Saves the most time of any prompt in this list once you save it as a template.

Read the email below. Then write 3 reply options:

1. Quick (1-2 sentences, friendly)
2. Direct (states the decision, no fluff, under 80 words)
3. Diplomatic (defers or negotiates, acknowledges their position first)

Match my usual voice: short sentences, em-dashes, no exclamation marks, no "hope this helps." Sign off with just my first name.

Email:
[PASTE EMAIL]

11. The "Counter-Argument" Prompt

For stress-testing your own thinking. The model will happily agree with whatever you say — unless you force it to disagree.

I'm about to commit to this decision: [STATE YOUR PLAN].

Steel-man the case AGAINST it. Don't be polite. Don't balance it with positives. Write as if you're a board member who thinks this is a mistake and I have 5 minutes to convince you otherwise.

List the 5 strongest reasons to abandon this plan, ranked. For each one, name the assumption of mine that has to be wrong for the reason to hold.

12. The "Prompt, Then Refine" Iteration Prompt

The meta-prompt: when you don't know what you want yet, ask ChatGPT to help you specify it before it answers.

Before you answer my question, ask me 3 clarifying questions that would help you give a much sharper answer. Don't answer the original question yet.

Once I respond, ask up to 2 more questions if needed, then deliver the final answer in the format I specified below.

Format: [bullet list / table / 200 words / etc.]

Question: [YOUR QUESTION]

The 6 Mistakes That Quietly Kill ChatGPT Output

Even with CRISPE, these six patterns will sabotage your results. They show up in 80% of weak prompts we review.

Mistake 1: Asking the model to do three things at once

"Summarize this article, give me 5 tweet ideas, and write a LinkedIn post." ChatGPT will do all three, but each will be shallower than if you'd asked for one. Fix: either split into 3 prompts, or explicitly tell the model to handle them sequentially with a clear separator.

Mistake 2: No success criteria

"Write me a good email." Good by whose standard? For what outcome? Fix: always close with a one-sentence win condition. "An email that gets a meeting booked by Friday" beats "a good email" every time.

Mistake 3: Trusting the first response

The first answer is the average answer. The third or fourth answer — after a few targeted refinements — is usually 2-3x better. Fix: treat the first response as a draft, not a final.

Mistake 4: Hedging your way into mediocrity

"Maybe you could try to think about whether…" loses ~30% of the model's effort to the hedging. Fix: lead with verbs. "Write. List. Compare. Rank. Decide."

Mistake 5: Forgetting to specify the output format

Without format instructions, the model returns a wall of prose. Fix: always say how the answer should be shaped — bullet list, table, JSON, 200 words, 3 sections. The model is dramatically better at fitting a shape than inventing one.

Mistake 6: Ignoring the model's training cutoff

For anything time-sensitive (news, prices, recent events), ChatGPT will hallucinate. Fix: paste the source material into the prompt, or explicitly say: "If you don't know, say 'I don't know' — do not guess."

Advanced: Chain-of-Thought and Few-Shot — When to Use Them

Two techniques from OpenAI's own prompt engineering guide show up over and over. Here's when each one earns its place.

Chain-of-thought (CoT) — adding "think step by step" or "walk me through your reasoning" before the final answer. It measurably improves accuracy on math, logic, multi-step analysis, and planning. Use it for: decisions, calculations, complex instructions, debugging. Skip it for: simple lookups, formatting tasks, anything that doesn't involve reasoning.

Few-shot prompting — including 2-3 examples of the exact output you want inside the prompt. Use it for: matching a specific style, format, or tone; teaching the model a pattern; getting consistent outputs across many similar requests. Skip it for: one-off questions where the format is obvious.

The two stack. Few-shot examples that include reasoning ("here's an example of a thought-out answer") outperform plain few-shot on hard problems — this is the variant from the original chain-of-thought paper by Wei et al.

Putting It All Together

You don't need to memorize CRISPE. You need to remember three questions before you hit enter:

1. Who should the model be? (Role)
2. What shape should the answer take? (Format)
3. What does "winning" look like? (Success criterion)

If you can answer all three in one sentence each, you'll outperform 90% of ChatGPT users. If you want the full six-part structure, save the CRISPE template in PromptLab's Builder and reuse it for every new prompt — that's what it's built for.

And once you've drafted a prompt with CRISPE, test it against your old version using PromptLab's Compare feature. Run both on ChatGPT, score them, see the difference. The 10x isn't theoretical — it's measurable.

Sebagian besar orang mengetik dua kalimat ke ChatGPT, dapat jawaban biasa-biasa saja, lalu menyalahkan AI-nya. Padahal modelnya bukan masalahnya — prompt-nya yang salah. Perbedaan antara jawaban datar dan jawaban tajam hampir selalu struktural, dan begitu lo pelajari strukturnya, model gratis yang kasih lo draft C- tadi akan kasih lo draft A dalam 4 detik.

Panduan ini adalah versi praktis, tanpa basa-basi, dari semua yang diajarkan dokumentasi resmi OpenAI — dikompres jadi satu framework, 8 template siap copy-paste, dan 5 kesalahan yang sering bunuh kualitas output. Kalau lo pernah bertanya cara buat prompt ChatGPT yang beneran deliver, ini jawabannya.

Kenapa Prompt Bahasa Indonesia Sering Gagal

ChatGPT dilatih dengan data bahasa Inggris 10× lebih banyak dari bahasa Indonesia. Artinya: kalau lo nulis prompt dalam bahasa Indonesia tanpa instruksi eksplisit, model akan nge-default ke respons generik, ambigu, dan terlalu aman. Modelnya bukan tidak bisa bahasa Indonesia — dia hanya tidak punya banyak "sinyal" bahasa Indonesia yang halus dan spesifik.

Solusinya: kasih lo role yang jelas, request yang spesifik, dan format output yang eksplisit. Lewat cara ini, model akan "berpindah mode" dari mode default ke mode profesional — dalam bahasa apa pun.

Framework CRISPE: Resep Prompt 6 Bagian

Framework paling konsisten yang sudah kami tes di ribuan prompt adalah CRISPE — singkatan dari Capacity, Request, Insight, Statement, Personality, Experiment. Yang penting: lo gak harus pake semua enam, tiga pertama saja sudah menutupi 80% kasus.

1. Capacity & Role

Kasih tahu ChatGPT dia siapa. Role menentukan level expertise, vocabulary, dan frame jawaban. Tanpa role, model default ke "asisten yang helpful" — versi paling low-effort dari dirinya.

Kamu adalah seorang copywriter senior yang spesialis menulis landing page untuk SaaS B2B. Gaya bahasamu punchy, spesifik, tanpa basa-basi — seperti gaya Harry Dry (MarketingExamples).

2. Request (tugas utamanya)

Langsung ke permintaan. Hindari "tolong bisa tolong" — itu cuma buang token dan gak nambah nilai.

Tulis 5 headline hero section untuk tool prompt engineering bernama PromptLab. Tiap headline maksimal 8 kata, fokus ke masalah (prompt jelek = output AI jelek), dan hindari kata "revolusioner", "powerful", "ultimate".

3. Insight (konteks latar belakang)

Semua info yang model butuhkan tapi gak ada di request: audiens, produk, tahap, apa yang sudah dicoba, apa yang dipertaruhkan.

Konteks: PromptLab adalah workspace prompt AI gratis. Target audiens: marketer dan content creator Indonesia yang sudah pakai ChatGPT harian tapi hasilnya inkonsisten. Diferensiasi utama: Readiness Score yang menilai struktur prompt. Kompetitor (Notion AI, Copy.ai) terlalu generik — kita mau yang lebih tajam dan punya opini.

4. Statement (kriteria sukses)

Ini bagian yang paling sering di-skip — dan ini kalimat dengan leverage paling tinggi di seluruh prompt lo.

Target: headline yang, ketika dibaca sendiri, bikin marketer berhenti scroll dan klik. Optimasi untuk kejelasan dan curiosity, bukan kepintaran.

5. Personality (nada, suara, gaya)

Jangan kasih adjective ("professional tapi hangat") — kasih contoh nama atau kalimat acuan.

Nada: kalimat pendek. Boleh pakai em-dash. TIDAK BOLEH pakai tanda seru. Mulai dengan verb atau masalah, jangan dengan nama brand. Gaya: Alex Hormozi bertemu Raditya Dika.

6. Experiment (format + iterasi)

Format output harus eksplisit: bullet list? tabel? 200 kata? Dan kalau model gak tau, harus tebak, tanya, atau tolak?

Format: return sebagai numbered list. Tiap item headline di baris sendiri, tanpa komentar tambahan. Kalau lo butuh lebih dari 8 kata, berhenti dan tanya dulu.

8 Template Prompt Siap Pakai (Copy-Paste)

Curang dari 12 template populer, ini 8 yang paling sering kepake untuk konteks Indonesia. Edit bagian dalam kurung siku sesuai kebutuhan lo.

1. Prompt "Berpikir Selangkah demi Selangkah"

Untuk analisis, planning, dan keputusan. Menambahkan instruksi "berpikir selangkah demi selangkah" meningkatkan akurasi di masalah multi-langkah — salah satu temuan paling konsisten di riset prompt engineering.

Kamu adalah konsultan strategi senior. Saya sedang mempertimbangkan [KEPUTUSAN].

Sebelum kasih rekomendasi, jalanin reasoning ini dulu:
1. Sebutkan 3 argumen TERKUAT untuk
2. Sebutkan 3 argumen TERKUAT kontra
3. Identifikasi 2 unknown-unknown yang bisa mengubah keputusan
4. Kasih rekomendasi akhir dalam 1 kalimat, dengan confidence level (rendah/sedang/tinggi)

2. Prompt Few-Shot (Tunjukkan, Jangan Ceritakan)

2-3 contoh lebih kuat dari 100 kata instruksi. Tepat untuk gaya, format, atau tone yang konsisten.

Samakan nada dan struktur 2 contoh tweet ini:

Contoh 1: Berhenti prompting kayak search engine. Mulai prompting kayak brief.
Contoh 2: AI lo cuma secerdas pertanyaannya. Kebanyakan pertanyaan itu ambigu.

Sekarang tulis 5 tweet lagi tentang [TOPIK] dengan suara yang sama. Tiap tweet max 200 karakter. Tanpa hashtag. Tanpa emoji.

3. Prompt "Bertindak Sebagai" Ahli

Klasik tapi efektif — pake ini kalau lo butuh model nge-draw dari vocabulary dan frame domain tertentu.

Bertindaklah sebagai akuntan pajak senior dengan 15 tahun pengalaman ngasih advice ke freelancer di Indonesia.

Pertanyaan saya: [PERTANYAAN]

Format jawaban: bullet point, 200 kata, sertakan referensi ke UU HPP atau PP 55/2022 kalau relevan.

4. Prompt Email Profesional Bahasa Indonesia

Email profesional Indonesia sering terlalu bertele-tele. Pakai ini untuk strip down ke esensi.

Baca email ini. Tulis 3 opsi balasan:

1. Singkat (1-2 kalimat, sopan)
2. Langsung (state keputusan, no fluff, max 80 kata)
3. Diplomatik (defer atau negosiasi, acknowledge posisi mereka dulu)

Cocokin nada profesional Indonesia yang umum di kantor: sopan, tidak terlalu formal, tidak pakai "Dengan hormat" atau "Hormat saya". Tutup dengan nama depan saja.

Email:
[TEMPEL EMAIL]

5. Prompt Counter-Argument

Untuk stress-test keputusan lo sendiri. Model biasanya cuma menyetujui — kecuali lo paksa dia untuk disagree.

Saya akan commit ke keputusan ini: [STATE RENCANA].

Buat kasus KONTRA yang terkuat. Jangan sopan. Jangan imbangi dengan positif. Tulis seolah lo adalah board member yang yakin ini salah dan saya punya 5 menit untuk meyakinkan lo.

List 5 alasan terkuat untuk meninggalkan rencana ini, diurutkan. Tiap alasan, tunjukkan asumsi saya yang harus salah supaya alasan itu berlaku.

6. Prompt Iterasi (Prompt, Lalu Perdalam)

Meta-prompt: kalau lo belum tau persis apa yang lo mau, minta ChatGPT nanya dulu sebelum jawab.

Sebelum jawab pertanyaan saya, tanya 3 pertanyaan klarifikasi yang akan bikin jawaban lo jauh lebih tajam. Jangan jawab pertanyaan aslinya dulu.

Setelah saya jawab, tanya max 2 pertanyaan tambahan kalau perlu, lalu kasih jawaban final dengan format di bawah.

Format: [bullet list / tabel / 200 kata / dll]

Pertanyaan: [PERTANYAAN ANDA]

7. Prompt Rangkum Riset (Anti-Halucination)

Untuk sintesis dokumen, paper, atau transkrip. Penting untuk memastikan model gak ngarang.

Bacalah dokumen terlampir. Lalu rangkum jadi 5 poin utama. Tiap poin harus:
- Bisa berdiri sendiri sebagai kalimat utuh
- Disertai kutipan atau rujukan ke bagian spesifik dokumen
- Ditandai confidence level (rendah/sedang/tinggi) berdasarkan seberapa eksplisit dokumen mendukungnya

Kalau ada klaim di dokumen yang menurut lo meragukan, flag dengan "⚠️ Klaim ini perlu verifikasi".

Dokumen:
[TEMPEL DOKUMEN]

8. Prompt Review Code (untuk Developer)

Lebih struktural dari "tolong review code ini". Hasilnya bisa langsung di-apply ke PR.

Review code di bawah sebagai senior engineer dengan standar production. Cari:

1. Bug yang jelas (logic error, off-by-one, null handling)
2. Edge case yang gak ke-handle
3. Performance issue (N+1 query, memory leak, blocking call)
4. Security issue (SQL injection, XSS, hardcoded secret)

Format tiap issue: severity (blocker/major/minor), lokasi file, deskripsi 1 kalimat, saran fix. Kalau code-nya bagus, bilang saja. Jangan padding.

Code:
[TEMPEL CODE]

5 Kesalahan yang Sering Bunuh Kualitas Output

Meskipun lo udah pake CRISPE, 5 pola ini akan sabotase hasil lo. Muncul di 80% prompt yang kami review.

Kesalahan 1: Minta 3 Hal Sekaligus

"Rangkum artikel ini, kasih 5 ide tweet, dan tulis LinkedIn post." ChatGPT akan lakukan semua tiga, tapi masing-masing akan lebih dangkal dibanding kalau lo request satu-satu. Fix: pecah jadi 3 prompt, atau eksplisit bilang model untuk handle sequentially dengan pemisah yang jelas.

Kesalahan 2: Gak Ada Kriteria Sukses

"Tulis email yang bagus." Bagus menurut siapa? Untuk outcome apa? Fix: tutup selalu dengan kalimat win condition satu baris. "Email yang bikin meeting kebooking hari Jumat" mengalahkan "email yang bagus" setiap saat.

Kesalahan 3: Mempercayai Respons Pertama

Jawaban pertama = jawaban rata-rata. Jawaban ketiga atau keempat, setelah beberapa refinement yang terarah, biasanya 2-3× lebih baik. Fix: perlakukan respons pertama sebagai draft, bukan final.

Kesalahan 4: Hedging yang Bikin Medioker

"Mungkin kamu bisa coba考虑 apakah…" kehilangan ~30% effort model ke hedging. Fix: pimpin dengan verb imperatif. "Tulis. List. Bandingkan. Rank. Putuskan."

Kesalahan 5: Lupa Training Cutoff

Untuk hal-hal yang sensitif terhadap waktu (berita, harga, event terbaru), ChatGPT akan halucinate. Fix: tempel material sumber ke prompt, atau eksplisit bilang: "Kalau gak tau, bilang 'saya gak tau' — jangan tebak."

Penutup: Mulai dari 3 Pertanyaan Ini

Lo gak perlu hafal CRISPE. Lo cuma perlu ingat 3 pertanyaan sebelum pencuk enter:

1. Siapa modelnya seharusnya? (Role)
2. Bentuk apa jawabannya? (Format)
3. Seperti apa "menang"-nya? (Kriteria sukses)

Kalau lo bisa jawab ketiganya dalam satu kalimat masing-masing, lo akan outperform 90% pengguna ChatGPT berbahasa Indonesia. Kalau mau struktur 6 bagian lengkap, save template CRISPE di PromptLab Builder — itu gunanya tools itu.

Dan setelah lo draft prompt dengan CRISPE, test melawan versi lama lo pake PromptLab Compare. Run keduanya, score, lihat perbedaannya. 10× bukan teori — bisa diukur.

Most marketing prompts produce copy that sounds like copy. The cadence is right, the format is right, and the conversion is still zero. The model isn't the problem — the prompt is. The difference between an AI paragraph that gets scrolled past and one that actually moves a metric comes down to a small set of structural decisions you make in the first 50 words of the prompt.

This guide is the practical, no-fluff version of what actually works in 2026 for marketers using AI. It contains 7 copy-paste templates — for social media, email, ads, landing pages, and product copy — each one tested against ChatGPT, Claude, and Gemini, and each one built on the same 6-part structure that our general prompt engineering guide recommends. If you've ever wondered how to create the best prompt for marketing that actually converts, this is for you.

Why Most Marketing Prompts Underperform

Marketing is a high-volume discipline. Most teams ship more copy in a week than a novelist ships in a year. That volume creates a temptation: keep the prompts short so you can move fast. The result is predictable — generic ad copy, on-brand-sounding captions that don't differentiate, and email subject lines that could be from any company in any industry.

The OpenAI prompt engineering guide, Anthropic's prompt engineering for business performance playbook, and the Microsoft prompt engineering documentation all converge on the same diagnosis: a marketing prompt needs more structure, not less. The most cited common failure is leaving the model to guess three things it can't possibly know: who the audience is, what conversion looks like, and what the offer is. When those three are missing, the model defaults to the statistical middle of its training data — which is the same place every other brand's copy is coming from.

A good marketing prompt is the opposite of generic. It names the funnel stage, the offer, the objection being addressed, and the conversion event. It tells the model to write as a specific persona, not as "a marketer." And it specifies the format, the length, and the call-to-action down to the verb.

The CRISPE-Marketing Framework: A 6-Part Prompt Recipe

The framework below is a marketing-tuned version of CRISPE — the 6-part prompt structure we cover in our Claude prompting guide and our general prompt engineering series. It is the same skeleton used by teams at companies like Anthropic and HubSpot to build production-grade marketing prompts.

1. Capacity & Role

Tell the model who it is. A role sets the expertise level, vocabulary, and the default frame for the response. For marketing, the role you pick changes the output dramatically — a "senior B2B SaaS copywriter" and a "DTC copywriter" produce completely different email subject lines from the same brief.

You are a senior conversion copywriter who specializes in B2B SaaS. Your writing style is sharp, specific, and benefit-driven — think Alex Hormozi meets Harry Dry (MarketingExamples). You never use the words "revolutionary," "powerful," or "ultimate."

2. Request (the actual marketing task)

The verb-forward, unambiguous ask. Lead with the action. Skip the "could you please" — it costs tokens and adds nothing to the response.

Write 5 LinkedIn ad headlines for a prompt-engineering tool called PromptLab. Each headline must be under 70 characters, lead with the pain (writing bad prompts = bad AI output), and end with a curiosity hook. Avoid superlatives.

3. Insight (background, audience, offer)

Everything the model needs to know that isn't in the request itself: funnel stage, audience, offer, price point, competitor copy, and what was tried before.

Context: PromptLab is a free AI prompt workspace. Target audience: marketing managers at B2B SaaS companies with 10-200 employees who already use ChatGPT daily but get inconsistent results. Offer: free, no credit card. Main differentiator: a Readiness Score that scores any prompt on structure. Competitor copy from Notion AI is too generic — we want sharper, more opinionated. Funnel stage: top-of-funnel awareness, optimizing for click-through to a free signup page.

4. Statement (the success criteria)

What does "winning" look like? For marketing, this is almost always a measurable outcome — click-through, reply rate, conversion, signups. State it explicitly.

Goal: a headline that, when read alone in a LinkedIn feed, makes a marketing manager stop scrolling and click. Optimized for clarity and curiosity, not cleverness. Benchmark to beat: 1.8% CTR (current LinkedIn ad average for B2B SaaS).

5. Personality (tone, voice, style)

Name the voice with examples, not adjectives. "Professional but warm" is meaningless. "Short sentences. Em-dashes. No exclamation marks. Starts with the punchline" is a prompt.

Voice: short sentences (max 12 words). Em-dashes allowed. No exclamation marks. Starts with the verb or the pain, never the brand. Banned words: revolutionary, powerful, ultimate, unleash, supercharge, robust. Reference voice: the Punch magazine copy in Alex Hormozi's $100M Offers.

6. Experiment (format + iteration)

How should the answer be shaped? Markdown table? JSON? Numbered list? 200 words? And what should the model do if it doesn't know — should it guess, ask, or refuse?

Format: return as a numbered list. Each item is the headline on its own line, no commentary, no preamble. If you would need more than 70 characters, stop and shorten before submitting. If you would use a banned word, pick a different one.

7 Copy-Paste Marketing Prompts That Convert

Steal these. Customize the bracketed parts. Each one is built on the CRISPE-Marketing framework above and tested across real ChatGPT, Claude, and Gemini conversations.

1. The LinkedIn Ad Headline Prompt

LinkedIn ads reward specific, pain-first headlines. The B2B funnel is slow and skeptical, and superlatives kill trust. This prompt produces 10 variations ranked by likely CTR.

You are a senior B2B SaaS copywriter who specializes in LinkedIn ads.

Write 10 LinkedIn ad headlines for [PRODUCT]. Each headline:
- Max 70 characters
- Leads with the pain (not the brand)
- Ends with a curiosity hook
- Avoids: revolutionary, powerful, ultimate, unleash, supercharge

Context: [AUDIENCE, OFFER, FUNNEL STAGE]
Goal: a headline that beats 1.8% CTR on LinkedIn.
Format: numbered list, one headline per line, no commentary.

2. The Email Subject Line Generator (With A/B Variants)

Email open rates are decided in the inbox preview pane. The subject line has 50 characters and one job: get the click. This prompt generates 12 subject lines across 4 psychological angles — curiosity, urgency, benefit, and personalization — so you can A/B test cleanly.

You are a senior email marketer. You write subject lines for a B2B newsletter with 50,000 subscribers and a 24% open rate baseline.

Generate 12 email subject lines for [CAMPAIGN]. Structure: 3 curiosity-based, 3 urgency-based, 3 benefit-driven, 3 personalized (use {{first_name}}).

Rules:
- Max 50 characters each
- Lowercase OK; no all-caps except for acronyms
- No clickbait ("You won't believe...")
- Each one should make a busy marketer stop and open

Format: return as a markdown table with columns: | # | Angle | Subject Line | Why It Works (1 sentence) |

3. The Landing Page Hero Section Prompt

The hero section is the highest-leverage 100 words on a landing page. This prompt builds the four parts of a high-converting hero — headline, subhead, primary CTA, secondary CTA — and asks for 3 variations so the design team can pick.

You are a senior conversion copywriter with 10 years of experience writing SaaS landing pages.

Write 3 variations of a hero section for [PRODUCT].

Each variation includes:
- Headline (max 8 words, punchy)
- Subheadline (max 25 words, explain the value, not the feature)
- Primary CTA button text (3-5 words, action verb)
- Secondary CTA text ("See How It Works" style)

Context: [AUDIENCE, OFFER, KEY DIFFERENTIATOR, COMPETITOR POSITIONING]
Goal: a hero that converts cold traffic at 4%+ (the 2026 median per Unbounce's benchmark report).
Tone: [professional / playful / authoritative — pick one]

Format: 3 clearly separated blocks. No explanation between them.

4. The Facebook / Instagram Ad Copy Prompt (Hook + Body + CTA)

Paid social ads fail when the hook doesn't match the body. The best Facebook ad copy in 2026 follows a tight 3-beat structure: scroll-stopping hook (first line), one-sentence benefit, single CTA. This prompt enforces that structure.

You are a direct-response copywriter who writes Facebook and Instagram ads.

Write 3 ad copy variations for [PRODUCT]. Each variation:

1. Hook (first line, max 40 characters, pattern-interrupt)
2. Body (max 125 characters, one benefit, one proof point)
3. CTA (single verb, max 5 words)

Audience: [WHO, what they care about, what objection they have]
Funnel stage: [cold / warm / retargeting]
Goal: a primary text that beats the 1.5% CTR benchmark for paid social.

Format: clearly labeled blocks. Banned phrases: "limited time," "don't miss out," "exclusive offer."

5. The SEO Blog Outline Prompt (Rank-Worthy Structure)

Most AI blog outlines are generic — 5 H2 sections of roughly equal weight, no search intent matched, no internal linking plan. This prompt produces an outline that respects Google's helpful content guidelines and matches the specific search intent.

You are an SEO content strategist who writes outlines that rank for competitive B2B keywords.

Create a detailed blog outline for: [TARGET KEYWORD]
Search intent: [informational / commercial / transactional]
Target audience: [WHO]
Word count target: [X] words

Include:
- Title (under 60 characters, includes the keyword)
- Meta description (max 155 characters)
- H1 (mirror the title, slightly more benefit-driven)
- 5-7 H2 sections, each addressing a distinct sub-question
- For each H2: 3-5 bullet points of what to cover
- An "FAQ" section at the end with 4 PAA-style questions (the "People Also Ask" box)
- 3 internal link anchors to existing [YOUR-SITE] content
- 1 external link to an authoritative source

Format: markdown outline, hierarchical.

6. The Product Description Prompt (E-commerce)

Product description copy on Shopify, Amazon, and DTC sites is repetitive, full of filler, and rarely differentiates. This prompt produces 3 variants of a product description optimized for scannability, emotional hook, and conversion.

You are a DTC copywriter who writes product descriptions for premium e-commerce brands.

Write 3 variants of a product description for [PRODUCT NAME].

Product features: [LIST]
Target buyer: [PERSONA, including one specific objection they have]
Price point: [BUDGET/MID/PREMIUM]

Each variant must include:
- Headline (max 8 words)
- 3 benefit bullets (each starting with a verb, not "Features include...")
- Emotional hook (1 sentence, sensory or aspirational)
- CTA (single verb, max 3 words)

Format: 3 clearly separated blocks. Avoid: "luxurious," "premium quality" (unless you can show why), "indulge."

7. The Customer Testimonial Repurposing Prompt

Customer quotes are the highest-trust marketing copy you have, and most teams waste them. This prompt takes one raw testimonial and pulls out the 4 most useful assets: a tweet, a LinkedIn post, an ad hook, and a headline pull-quote.

You are a content marketer who repurposes customer testimonials into high-converting micro-copy.

From the testimonial below, extract:
1. A pull-quote headline (max 8 words, preserve the customer's voice)
2. A tweet-length social post (max 240 characters, attribution at end)
3. A LinkedIn post (150-200 words, frame as a customer success story)
4. A Facebook ad hook (max 40 characters, pattern-interrupt style)

Rules:
- Do not invent claims. Use only what the customer said.
- Preserve the customer's phrasing when it is stronger than yours.
- For each asset, cite which sentence in the original you pulled from.

Format: clearly labeled, with the source sentence quoted for each.

Testimonial:
[PASTE HERE]

3 Marketing Prompt Mistakes That Kill Conversion

Even with the templates above, three patterns consistently undermine marketing output. Spotting them in your own prompts is the difference between copy that ships and copy that gets stuck in revision.

Mistake 1: Writing for "Everyone"

"Write copy for small business owners" produces copy that resonates with no one. Marketing copy is most effective when it speaks to one specific person, in a specific role, with a specific objection. The prompt above replaces "audience" with as much demographic and psychographic detail as you can fit in 2-3 sentences. If you can't fill in the bracketed parts of the templates, you don't know your audience well enough to write copy for them — and the model can't fill the gap.

Mistake 2: Skipping the Funnel Stage

Cold traffic and warm retargeting need different copy. Cold traffic needs an interruption — a hook, a question, a counterintuitive claim. Warm retargeting needs reinforcement — a feature, a comparison, a number that proves the claim. A prompt that doesn't specify the funnel stage produces average output for both. The templates above name the stage explicitly in the Context block. Don't skip it.

Mistake 3: Trusting the First Output

Marketing copy is the highest-volume use case for LLMs, and the first output is rarely the best. Treat the first response as a draft, then iterate with one targeted refinement. The most common high-leverage refinement is: "Rewrite the 3 strongest variations, but with a specific stat or proof point from this list: [paste proof points]." Adding a real number to a generic claim is the single fastest way to lift CTR on a marketing prompt.

Putting It All Together

The 7 templates above cover the channels most marketing teams ship copy into: LinkedIn ads, email, landing pages, paid social, blog content, product pages, and testimonial repurposing. The framework that ties them together is the same one we use across our other marketing prompt templates and our ChatGPT vs Claude comparison: name the role, state the request, give the context, define the success metric, specify the voice, and lock the format.

That's it. Seven templates, one framework, no magic words. The difference between marketing copy that gets ignored and marketing copy that converts is structural — and structure is a skill you can build in an afternoon.

Try these techniques in PromptLab. Paste any of the 7 templates into the builder, fill in the bracketed parts, and run it against GPT, Claude, and Gemini side-by-side. The Readiness Score tells you whether the prompt is structurally tight, and Compare shows you which model produces the most on-brand output. Marketing prompts are measurable — and the improvement is visible in the first 30 minutes.

Take the same prompt — “Write a 200-word LinkedIn ad for an AI prompt tool, pain-first, no superlatives” — and run it on Claude 4.7, GPT-5, and Gemini 2.5. You will get three different outputs. One will lead with the pain, one will lead with the brand, and one will produce a paragraph that reads like a press release. The model you pick changes the result more than the words you write.

This guide is the model-specific playbook for the three models that matter in 2026. It is built on the official prompting guidance from Anthropic, OpenAI, and Google — and tested against the same 5 production prompts in our Claude guide and our ChatGPT vs Claude comparison. By the end, you will know exactly which prompt pattern fits each model, and which one to reach for when you need a particular output shape.

Why One Prompt, Three Different Models, Three Different Answers

Every modern LLM is trained on the same internet, but they are not the same model. They differ in three ways that matter for prompting: pre-training objective (what they were optimized to be good at), post-training alignment (how RLHF tuned their voice and refusal behavior), and inference infrastructure (what reasoning modes the API exposes). The result is that the same prompt lands on three different “personalities.”

OpenAI’s official GPT-5 Prompting Guide says it directly: GPT-5 is “very adaptive and requires very specific instructions,” and was trained to be a strong foundation for agentic and tool-calling flows. Anthropic’s Claude Prompting Best Practices emphasizes XML-style structure, long-context handling, and adaptive thinking. Google’s Prompt Design Strategies for Gemini focuses on natural-language flexibility, multimodal inputs, and thinking models like Gemini 2.5 Pro.

The gap is not intelligence. All three are at the frontier. The gap is in shape: which instructions each model translates most reliably into the output you want. The rest of this guide is the shape-by-shape map.

Quick Comparison: Claude 4.7 vs GPT-5 vs Gemini 2.5 Pro

Before the deep dive, here is the 90-second view. Use this as a decision tree when you are picking a model for a specific task.

• Best for long, structured reasoning: Claude 4.7 Sonnet/Opus. XML tags, multi-step plans, and careful instruction-following.
• Best for agentic + tool calling: GPT-5. The Responses API persists reasoning across tool calls, and the model is tuned to plan-and-execute.
• Best for multimodal + real-time grounding: Gemini 2.5 Pro. Native image, video, and audio understanding, plus Google Search grounding in the API.
• Best for short, creative copy: GPT-5 — but only with explicit length and format constraints.
• Best for code review and careful refactors: Claude 4.7 Sonnet — more literal, less overeager, fewer false positives.
• Best for research synthesis across many sources: Gemini 2.5 Pro with thinking enabled — 1M+ token context, ground with Search.

How to Write the Best Prompt for Claude 4.7

Claude is the model to reach for when the prompt has many moving parts. It is the most literal instruction-follower at the frontier, which is a feature when you have a complex brief and a bug when your brief is sloppy. Anthropic’s official guidance puts it this way: “Be clear and direct. Add context to improve performance. Use examples effectively.”

1. Use XML Tags to Separate Sections

Anthropic explicitly recommends XML-style tags for structured prompts. The model is trained to respect XML boundaries, so instructions inside <instructions> are interpreted as instructions, not as data or context. The same prompt without tags will produce an output that “blurs” the sections — the model starts to follow an example as if it were an instruction.

<role>
You are a senior B2B SaaS copywriter with 10 years of experience.
You write in the style of Alex Hormozi and Harry Dry.
</role>

<context>
Product: PromptLab, a free AI prompt workspace.
Audience: marketing managers at B2B SaaS companies, 10-200 employees.
Funnel stage: top-of-funnel awareness. Offer: free, no credit card.
</context>

<task>
Write 5 LinkedIn ad headlines. Each one:
- Under 70 characters
- Leads with the pain, not the brand
- Ends with a curiosity hook
- Banned words: revolutionary, powerful, ultimate, unleash
</task>

<output_format>
Return as a numbered list. One headline per line. No commentary.
</output_format>

2. Turn On Adaptive Thinking for Hard Tasks

Claude 4.7 ships with adaptive thinking — instead of a fixed reasoning budget, the model decides how much to think based on the prompt. For analytical, multi-step, or strategic work, set the effort parameter high and let the model reason out loud before it answers. For simple lookups and short copy, set effort low or off — you want speed, not analysis paralysis.

Anthropic’s docs call out a specific 4.6+ migration: adaptive thinking replaces the older budget_tokens parameter. If you are upgrading an existing app, swap the budget for thinking: {type: "adaptive"} and tune the effort field.

3. Use Prefill and Stop Sequences to Lock the Output Shape

Claude supports prefill — you can write the first few tokens of the assistant’s response before the model takes over. The Anthropic prompt engineering guide recommends prefill as a way to force JSON, skip pleasantries, and bias the model toward a specific output structure. (As of Claude 4.6+, the API is moving away from prefill for some workflows in favor of explicit response preambles — check the migration notes if you have an older app.)

Assistant: {

When the assistant’s response must start with a JSON brace, the model will fill in the object instead of writing “Sure! Here is a JSON object:”. That single change often doubles the parsing reliability of downstream code.

4. Long Context: Put the Document First, Then the Question

Claude’s 200K+ token context window means you can paste entire books, codebases, or transcripts. The Anthropic guidance: put the long document first, then the instructions, then the question. This mirrors how the model is trained on long-context retrieval and reliably improves answer accuracy on “needle in a haystack” questions by 5-15%.

How to Write the Best Prompt for GPT-5

GPT-5 is the most “adaptive” model at the frontier. OpenAI’s official guide makes the point explicit: minimal prompts work best for Codex-style tasks, but for general-purpose prompts you should be precise and avoid conflicting information. The single biggest mistake people make with GPT-5 is treating it like GPT-4 — over-engineering system prompts and adding rules that conflict with each other.

1. Set the Right Reasoning Effort

GPT-5 introduces a reasoning_effort parameter with four settings: minimal, low, medium, and high. The default is medium. The OpenAI guide is explicit: for coding and agentic tasks, you usually want medium or high. For one-line answers, classification, or extraction, set minimal and the model will answer in a single forward pass with no chain-of-thought overhead.

The mistake: leaving reasoning_effort at the default medium for a quick lookup. The model will burn 3-10x more tokens than needed to reason about a question that does not need reasoning. Set it explicitly. The cost savings are real.

2. Use the Responses API for Agentic Work

OpenAI’s guide recommends the Responses API over the older Chat Completions API for any agentic or tool-calling flow. The reason: reasoning state is persisted between tool calls. The model can think about the result of one tool call, decide what to call next, and only emit a final answer when it is ready. With Chat Completions, you have to rebuild the reasoning state on every call — slower, more expensive, and prone to loops.

3. Tell GPT-5 to Proceed Under Uncertainty

One of the highest-leverage prompt patterns from the GPT-5 guide: tell the model to proceed when it is uncertain instead of stopping and asking. The example from the docs: “Never stop or hand back to the user when you encounter uncertainty. Decide what the most reasonable assumption is, proceed with it, and document it for the user’s reference after you finish acting.”

This is the opposite of the Claude default, which is more conservative and more likely to ask. For agentic workflows that need to ship, GPT-5’s bias toward action is a feature — but only if you tell the model to document its assumptions as it goes, so you can audit them.

4. Be Specific About Tool Use

GPT-5 calls fewer tools by default than 3.5. If you want the model to use web search, file search, or a custom connector, you have to say so explicitly. The OpenAI guide recommends phrasing tools as optional but encouraged rather than mandatory: “Use web search when the answer depends on information that may have changed since the training cutoff. Cite at least one source per non-trivial claim.”

You are a research analyst. For the question below, use web search when the answer depends on recent data. Cite sources inline with [1], [2] notation. If you use a tool, summarize what you searched for and what you found before giving the final answer.

Question: [YOUR QUESTION HERE]

How to Write the Best Prompt for Gemini 2.5 Pro

Gemini is the most flexible of the three — it accepts natural-language prompts in almost any shape and is trained to be the most conversational. The Google Prompt Design Strategies documentation emphasizes that Gemini is designed to be talked to like a person, and the model’s biggest strength is multimodal: it can read images, video frames, and audio alongside text in a single prompt.

1. Lean Into Multimodal Inputs

If your task involves a screenshot, a photo, a chart, or a video, Gemini is the model to reach for first. The Google blog introducing Gemini 2.5 Pro emphasizes the model’s “enhanced performance and improved accuracy” on multimodal reasoning, and the API accepts images, video frames, and audio natively without any preprocessing.

For text-only tasks, Gemini is still strong — but the differentiator is the multimodal angle. The most underrated use case is uploading a screenshot of a UI bug and asking for a fix: Gemini reads the layout, identifies the CSS class, and proposes a patch in a single shot.

2. Use Thinking Mode for Multi-Step Reasoning

Gemini 2.5 Pro is a thinking model — it reasons before it answers. The Google docs recommend explicitly turning thinking on for math, code, and research tasks. The model will show its work (configurable), and the answer quality is materially better on multi-step problems than the non-thinking Gemini 2.0 Flash baseline.

The pattern: combine thinking with structured output. Tell the model to reason step-by-step, then emit a final JSON object. Gemini handles the format better than Claude or GPT-5 when the format is JSON Schema — the model’s training data includes more schema-typed examples.

3. Ground with Google Search

For any task that depends on current information — pricing, recent news, stock prices, latest product features — use Gemini with the Search tool enabled. The Google AI Studio lets you toggle Search grounding with one click, and the API exposes it as tools: [{google_search: {}}]. The model will cite sources inline and link to the pages it used.

This is Gemini’s killer feature for research workflows. The combination of 1M+ token context, multimodal input, and real-time Search grounding means a single Gemini call can replace a multi-step research pipeline that would otherwise take 5-10 Claude or GPT-5 calls.

4. Describe the Scene, Don’t List Keywords

For image generation with Gemini 2.5 Flash Image, the Google Developers Blog is explicit: “Describe the scene, don’t just list keywords.” The model’s core strength is language understanding, so a paragraph describing the composition, lighting, and mood produces a much better image than a comma-separated list of style tags. This is the opposite of how Midjourney is tuned — see our image generation guide for the side-by-side comparison.

Side-by-Side: The Same Prompt, Three Models

Here is a real test we ran on all three models. Same prompt, same model parameters, no post-editing. The differences are immediate and instructive.

The Test Prompt

You are a senior conversion copywriter. Write 3 LinkedIn ad headlines for an AI prompt tool called PromptLab. Each headline must be under 70 characters, lead with the pain (not the brand), and end with a curiosity hook. No superlatives.

Claude 4.7 Sonnet Output

1. Your prompts are embarrassing the model.
2. Stop asking ChatGPT to “write me a thing.”
3. Your AI copy sounds like everyone else’s.

Verdict: Leads with the pain in every line, sharp and specific. The tone is the closest to a senior copywriter — opinionated, not generic. The headlines feel like a writer with taste wrote them, not a model averaging the internet.

GPT-5 Output (reasoning_effort: medium)

1. Bad prompts, bad output. There is a better way.
2. Your AI gives you 3 drafts. You rewrite all 3.
3. Why your ChatGPT results feel generic (and the fix).

Verdict: Cleaner structure, more obviously “ad-shaped.” Each headline has a number or a specific pattern (parenthetical, em-dash). The tone is professional and on-brand, but slightly less sharp than Claude’s. GPT-5 biased toward clarity and pattern-matching, which works well for B2B.

Gemini 2.5 Pro Output (with thinking)

1. Writing prompts is a skill. Here is how to learn it.
2. The 3 lines that turn ChatGPT from a toy into a tool.
3. Your AI is only as good as what you ask it.

Verdict: More conversational, slightly more generic. The third headline is a cliche. Gemini leaned toward the “listicle hook” pattern, which works for educational content but feels softer than Claude or GPT-5 for direct-response. Strongest output would be the second headline — the “3 lines that turn” pattern is unique.

One Unified Template That Works on All Three

If you have to write one prompt that runs on any of the three models — a common situation in production apps that route to different providers for cost or latency — use this template. It is built on the 5-part CRISPE structure from our general prompt engineering guide, and it is tuned to avoid the model-specific failure modes above.

<role>
You are a [SPECIFIC ROLE: title, years of experience, domain].
Voice reference: [NAME 2-3 WRITERS OR BRANDS TO MIMIC].
Avoid: [LIST 3-5 BANNED WORDS OR PHRASES].
</role>

<context>
Audience: [JOB TITLE, COMPANY SIZE, MAIN OBJECTION].
Offer: [WHAT YOU ARE SELLING, PRICE POINT, DIFFERENTIATOR].
Funnel stage: [AWARENESS / CONSIDERATION / RETARGETING].
</context>

<task>
[VERB-LED ACTION]: "Write 3 LinkedIn ad headlines" / "Audit the landing page" / "Generate 5 subject lines".
[QUANTITY AND CONSTRAINTS]: length, count, format, banned words.
[SUCCESS CRITERIA]: a specific metric or behavioral outcome.
</task>

<output_format>
Return as [numbered list / markdown table / JSON object / document].
[SPECIFIC SHAPE: 3 columns, 70 chars max, etc.].
No preamble. No recap. No filler.
</output_format>

<rules>
- Do not use [BANNED WORD 1], [BANNED WORD 2], or [BANNED WORD 3].
- Every line must lead with the pain, not the brand.
- If you would need more than [X] characters, stop and shorten.
</rules>

This template works on all three models because it removes the model-specific failure modes. It uses XML tags (which Claude respects), explicit reasoning (which GPT-5 prefers), and a structured output shape (which Gemini handles well). The Rules block is a “do this” version of every “don’t” — which Anthropic specifically recommends in their latest guidance.

3 Model-Specific Mistakes to Avoid

Even with a great template, three patterns consistently produce bad output on the wrong model. Spot them before you ship.

Mistake 1: Asking Claude to Reason Without Giving It a Structure

Claude 4.7 with adaptive thinking will reason — but it reasons best when given a structure to reason within. A prompt that says “think about this carefully and give me your best answer” produces a longer, more cautious response than a prompt that says “list the 3 strongest arguments for X, then pick one and defend it.” Claude’s adaptive thinking allocates effort based on the structure of the request. Vague structure = vague reasoning.

Mistake 2: Asking GPT-5 to Do Quick Lookups With Reasoning Turned On

The OpenAI guide is explicit: “For one-line answers, set reasoning_effort = minimal.” A lookup question with medium reasoning is 3-10x more expensive and 2-5x slower than the same call with minimal. The output quality is essentially identical for short answers. The default medium is the right call for complex tasks and the wrong call for everything else.

Mistake 3: Forgetting to Ground Gemini for Real-Time Tasks

Gemini 2.5 Pro has a knowledge cutoff. If you ask it for the current price of a stock, the latest version of a software product, or a recent news event, the model will answer from training data — which may be months or years out of date. The fix is one line in the prompt: “Use web search to find the current value. Cite the source.” Without that line, you will get confident answers that are subtly wrong.

Putting It All Together

The three models are not interchangeable. They are three different tools built for three different jobs. The job of the prompt engineer is to know which model fits which job, and to write the prompt in the shape that model handles best. Use Claude for careful, structured, long-form work. Use GPT-5 for agentic, tool-using, fast-iterating workflows. Use Gemini for multimodal, real-time, and Search-grounded tasks. For everything else, the unified template above is a solid default.

That is the playbook. One framework, three models, no magic words. The model you pick is the biggest variable in your output — and the prompt is the second biggest. Master both, and the difference between mediocre and great AI output is no longer a mystery.

Try these techniques in PromptLab. Paste the unified template into the builder, run it against Claude 4.7, GPT-5, and Gemini 2.5 Pro side-by-side, and the Readiness Score will tell you which prompt structure is the tightest. Then save the winning variant as a reusable template. Model selection is measurable — and the improvement shows up in the first session.

Type "a cat" into Midjourney, DALL-E, and Flux and you get three different cats in three different styles on three different backgrounds. Type a 6-part structured prompt and you get a portfolio-grade render that looks the way you actually imagined it. The gap between those two outputs is not the model — it is the prompt. This guide is the 6-part image prompt formula that closes that gap, with model-specific tuning for the three generators that matter in 2026.

Everything below is built on the official documentation from our general prompt engineering guide, Midjourney's Prompt Basics, the FLUX.2 Prompting Guide from Black Forest Labs, and OpenAI's DALL-E guidance. If you have ever pasted a one-line idea and gotten a one-line image back, the rest of this article is the fix. For a side-by-side comparison of how the same prompt lands on ChatGPT vs Claude vs Gemini, see our Claude vs GPT-5 vs Gemini breakdown — image generation is a separate but related skill, and the same model-specific thinking applies.

Why "A Cat" Is Not a Prompt

Image generators are not search engines. They do not retrieve "a cat" — they synthesize a cat from noise, conditioned on every word in your prompt. A vague prompt hands the model 95% of the creative decisions. A structured prompt hands the model exactly the 4-6 decisions that matter and lets it spend the rest of its capacity on rendering quality. The result of the second kind of prompt is sharper, more on-brief, and closer to what you actually pictured.

Midjourney's own docs make the point sharper: "Short and simple prompts typically generate the best images with Midjourney." The trick is that "short" does not mean "vague." It means "every word does work." The FLUX.2 Prompting Guide agrees: priority order is main subject → key action → critical style → essential context → secondary details. Every word in a good image prompt is on that priority list.

Black Forest Labs puts a hard cap on it: the FLUX.2 prompt length guidance is 2-3 sentences, with the most important elements first. OpenAI's DALL-E guidance, by contrast, tolerates (and rewards) longer natural-language descriptions. The three models want the same information, but they want it in three different shapes. That is the whole game.

The 6-Part Image Prompt Formula

Before we go model-by-model, here is the universal formula. It works on Midjourney, DALL-E, and Flux (with the model-specific tweaks in the next sections). Use it as a checklist whenever you sit down to write a prompt — most of the time, the difference between a weak output and a strong one is whether you filled in all 6 parts.

1. Subject — the main object, person, animal, or scene. Be specific: "a calico cat" beats "a cat". "three children in raincoats" beats "kids."
2. Action / Pose — what the subject is doing. "mid-leap," "looking over its shoulder," "hands in pockets." If you skip this, the model guesses and the guess is usually "standing still and looking at the camera."
3. Environment / Setting — where the subject is. "on a wet cobblestone street at dusk," "in a glass terrarium," "floating in zero gravity."
4. Lighting — how the scene is lit. "soft window light from camera left," "harsh neon underlight," "golden hour backlight." Lighting is the single biggest factor in mood and the easiest one to skip.
5. Style / Medium — the artistic rendering. "shot on 35mm film," "watercolor illustration," "3D isometric render in Blender," "Studio Ghibli animation cel."
6. Composition / Framing — where the camera is. "wide shot, low angle, rule of thirds," "extreme close-up of hands," "top-down flat lay."

Here is the same 6-part formula applied to a real prompt. Read it left-to-right and notice how every part has a job:

A calico cat (Subject)
mid-leap off a wooden fence (Action)
in a sunlit cottage garden full of lavender (Setting)
lit by warm golden-hour backlight (Lighting)
shot on 35mm film with shallow depth of field (Style + Composition)
--ar 3:2 --v 7 (Midjourney parameters)

That single prompt would land well on all three models. The next three sections show you how to tune it for each one specifically.

How to Write the Best Midjourney Prompt

Midjourney is tuned for short, evocative, comma-separated prompts with a clear subject, medium, and mood. The official Prompt Basics doc is explicit: "Short and simple prompts typically generate the best images." But "short" means every word pulls weight, not that the prompt is underspecified. The trick is to think in modifiers, not sentences.

1. Lead with the Subject and Action

Midjourney weighs the first few words the most. Put the subject and what it is doing at the front, then layer on style and environment. The model treats modifiers as ranking cues — earlier words dominate the output, later words adjust.

A weathered sailor in a yellow slicker, steering a fishing boat through a storm, cinematic photograph, dramatic lighting, ocean spray, shot on 35mm film, --ar 16:9 --v 7 --style raw

The --style raw parameter reduces Midjourney's default beautification and gets you closer to the prompt you actually wrote. Use it for product shots, technical illustrations, and any time the default looks "too polished."

2. Use the --no Parameter Instead of "No" Words

The Midjourney docs are clear: "Describe what you do want instead of what you don't. If you mention a party with 'no cake,' a cake might still appear." Negations are unreliable in diffusion models. Use the --no parameter to exclude elements you do not want in the output.

Editorial portrait of a female CEO in a minimalist office, soft window light, 85mm lens, neutral expression, no smile --no smile teeth makeup

The --no parameter is a hard exclusion — Midjourney actively steers the image away from those tokens. The list should be short (3-5 terms) and specific. Do not try to exclude abstract concepts like "boring" or "generic." Stick to concrete visual elements.

3. Stack Style References with --sref for Visual Consistency

Midjourney's --sref (style reference) parameter lets you anchor a generation to the visual language of another image. For brand work, character sheets, and product families, this is the most underrated tool in the suite. Pick a single reference image whose style you want to imitate, upload it, and pass its URL with --sref URL.

A minimalist product hero shot of a ceramic coffee mug on a marble surface, soft natural light, lifestyle photography --sref https://example.com/your-brand-style.jpg --ar 4:5 --v 7

For a series, use the same --sref URL across 20-30 prompts and the output will hold a consistent visual language — same color grade, same lighting feel, same level of polish. This is the closest thing Midjourney offers to a "brand kit," and it is the highest-leverage tool for any work that ships in volumes.

4. Use Multi-Prompts and Weights for Fine Control

Midjourney's :: syntax lets you weight parts of a prompt. A part with ::2 is twice as influential as a part with no weight. This is the right tool when one element needs to dominate the output and another needs to be a background detail.

red fox in a snowy forest::2 autumn leaves on the ground::1 soft bokeh background::0.5 --ar 3:2

The first part (red fox) is the dominant subject, the second (autumn leaves) is mid-priority, the third (bokeh) is a subtle background modifier. Read the weights as a hierarchy, not as a recipe.

How to Write the Best DALL-E Prompt

DALL-E is the most "conversational" of the three image generators. It accepts natural-language paragraphs and tends to follow longer, more descriptive prompts better than Midjourney or Flux. OpenAI's guidance: be specific about the scene, the lighting, the camera, the mood, and the style, and let the model interpret the prose.

1. Write a Scene Description, Not a Tag List

The DALL-E 3 prompting community has converged on a clear pattern: paragraphs beat tag lists. A scene description reads like a film director's shot list — subject, action, environment, lighting, camera. The model interprets the prose, infers the relationships, and produces a coherent image.

A cinematic wide shot of a lone astronaut standing on the edge of a red Martian canyon at sunset. The astronaut's visor reflects the canyon walls. The sky is dusty pink and amber. Low camera angle, looking up at the astronaut. Photorealistic, shot on ARRI Alexa, anamorphic lens flare.

Notice the structure: opening framing (cinematic wide shot), subject (astronaut), setting (Martian canyon at sunset), specific detail (visor reflection), color (dusty pink, amber), composition (low angle, looking up), rendering (photorealistic, ARRI Alexa). DALL-E 3 handles this kind of paragraph better than a comma-separated tag list.

2. Specify Style and Reference Artists Carefully

DALL-E will produce images in the style of named artists, art movements, and even specific films — but the results depend heavily on the model's training data. Photography styles (35mm, polaroid, shot on iPhone, National Geographic) work very reliably. Art movements (Bauhaus, art deco, impressionism) work well. Named living artists are often rejected or produce inconsistent results. If you want a 1940s noir look, say so explicitly: "1940s film noir, high-contrast black and white, venetian blind shadows, femme fatale silhouette." That works. "Like that movie" does not.

3. Avoid "Don't" — Use "Do" Instead

Like Midjourney, DALL-E handles negations unreliably. Saying "no text" or "no watermark" often produces text and watermarks. The fix is to say what you want instead of what you do not want, and to make the positive instruction as concrete as possible.

A clean product photograph of a white ceramic vase on a seamless light gray background, soft even studio lighting, no visible text, no logos, no props, no reflections

Notice the trick: the prompt says "no text, no logos, no props, no reflections" — but those words are surrounded by positive framing ("clean product photograph," "seamless light gray background") that anchors the model in the do-state. Pure negation produces unpredictable results; positive framing with selective negative qualifiers is much more reliable.

4. For Text in Images, Use Quotation Marks and Specify the Font

One of the most-requested DALL-E features is reliable text rendering. The community guidance: use quotation marks around the literal text, specify a font style, and keep the text short (3-5 words is the sweet spot).

A vintage travel poster for Paris, the text "PARIS" in large bold art-deco serif letters at the top, "BONJOUR" in smaller letters below, soft watercolor background of the Eiffel Tower at sunset

For longer text, DALL-E still struggles. If you need a poster with a paragraph of body copy, the right workflow is to render the image without text, then composite the text in a tool like Figma or Photoshop. Treating DALL-E as the image layer and a design tool as the typography layer is the reliable production workflow.

How to Write the Best Flux Prompt

FLUX (Black Forest Labs) is the newest of the three, and its prompt format reflects that. The official FLUX.2 Prompting Guide is explicit: "No negative prompts. FLUX.2 does not support negative prompts." And: "FLUX.2 generates photorealistic images from simple, natural language prompts." The format is closer to DALL-E (natural language paragraphs) than to Midjourney (tag lists with parameters).

1. Use the Subject + Action + Style + Context Framework

Black Forest Labs' official guidance for FLUX.2 is a clean four-part framework: Subject + Action + Style + Context. Priority order is main subject → key action → critical style → essential context → secondary details. Put the most important elements at the start of the prompt; let the modifiers fall where they may.

A determined young woman (Subject) climbing a sheer rock face (Action), in cinematic wide-angle style with golden hour backlighting (Style), in a misty mountain valley at sunrise (Context)

The four parts are easy to write quickly, and the priority order is the same as the priority order of attention in the model. The result is a prompt that the model can render with high fidelity on the first try.

2. Use Hex Codes for Brand Colors

One of FLUX.2's most underrated features: it understands hex codes. The official docs are explicit: "Use hex codes for brand text: 'The logo text "ACME" in color #FF5733'." For brand work — logos, packaging, marketing assets — this is the cleanest way to lock in a specific color.

A minimalist product render of a stainless steel water bottle on a pure white background, the brand text "HYDRO" in bold sans-serif letters in color #2A9D8F on the side of the bottle, soft studio lighting, no shadows

Compare that to the DALL-E version of the same prompt, which would have to describe the color as "teal green" or "muted blue-green" and would produce variable results. The hex code removes ambiguity.

3. Do Not Use Negative Prompts (Use Positive Framing)

FLUX.2 does not support the --no style negative prompt. To exclude something, describe the output you want in positive terms — and the model will not render the unwanted element if the positive description is strong enough. "A clean product photograph on a pure white background, no visible shadow, no other objects, no people" works because the positive framing ("clean," "pure white," "no shadow") anchors the model. The negations reinforce; they do not lead. Same approach as DALL-E, same reason it works.

4. For Photorealism, Specify Camera and Lens

FLUX.2 is trained heavily on photographic data, and the model's strongest photorealism comes from prompts that name the camera, lens, and lighting setup. The community has converged on a clean pattern: "shot on [camera] with [lens] under [lighting]".

A candid street photograph of an elderly man feeding pigeons in a European city square, shot on a Leica M11 with a 35mm Summicron lens, late afternoon natural light, shallow depth of field, grain

The named camera (Leica M11) and lens (35mm Summicron) trigger specific photographic training data in the model. The result is a more "photographic" image than a generic "candid street photograph" would produce. This is the single highest-leverage technique for FLUX.2 photorealism, and it composes with the next section: same prompt, three different formats, three different results.

Side-by-Side: The Same Prompt, Three Models

Here is a real test. Same idea, three models, three different prompt formats. The differences are immediate and instructive.

The Test Idea

A tired barista in a small independent coffee shop at 6am, golden hour light through the front window, steaming espresso in hand, candid documentary photography style.

Midjourney Output

Tired barista holding a steaming espresso, early morning, golden hour sunlight through window, documentary photography, 35mm film, shallow depth of field --ar 3:2 --v 7 --style raw

Verdict: Short, comma-separated, parameters at the end. Midjourney will produce a moody, film-grain image with strong directional light. The --style raw keeps it from looking "too AI." The short format is exactly what Midjourney is tuned for.

DALL-E Output

A candid documentary-style photograph of a tired barista in a small independent coffee shop at 6am, holding a steaming espresso. Golden hour light streams through the front window, catching the steam. The barista looks directly at the camera with a quiet, end-of-shift expression. Shot on 35mm film, shallow depth of field, warm tones.

Verdict: Natural-language paragraph with explicit scene direction. DALL-E will produce a more "narrative" image — the model interprets the paragraph as a story beat and renders the relationships (steam in light, barista's expression, the window direction). The format is what DALL-E is tuned for.

Flux Output

A tired barista in a small independent coffee shop at 6am, holding a steaming espresso. Golden hour light streams through the front window, catching the steam. The barista looks directly at the camera with a quiet, end-of-shift expression. Shot on a Leica M11 with a 35mm Summicron lens, warm tones, shallow depth of field.

Verdict: Almost identical to the DALL-E prompt, with one key swap: "shot on 35mm film" → "shot on a Leica M11 with a 35mm Summicron lens." That single change unlocks FLUX.2's strongest photorealism. The rest of the paragraph structure is the same — natural language, four-part framework, explicit lighting.

The Unified Template That Works on All Three

If you have to write one prompt that runs on any of the three models — a common situation in production apps that route to different providers for cost, latency, or capability — use this template. It is built on the 6-part formula from our Claude prompting guide (which applies to image prompts too — Claude and other LLMs are great for expanding a one-line idea into a structured image prompt), and it is tuned to avoid the model-specific failure modes above.

SUBJECT (1 line): [Main subject + key identifying detail]
ACTION (1 line): [What the subject is doing, in a specific verb]
ENVIRONMENT (1 line): [Where the subject is, with 1-2 sensory details]
LIGHTING (1 line): [Direction, quality, and color of the light]
STYLE (1 line): [Camera/lens, art movement, or rendering style]
COMPOSITION (1 line): [Framing, angle, depth of field]
PARAMETERS (optional, Midjourney only): --ar 16:9 --v 7 --style raw
NEGATIVE (optional, Midjourney only): --no [3-5 specific terms to exclude]

Fill in the six labeled lines, leave the last two for Midjourney. The result is a prompt that produces a strong output on all three models without rewriting for each.

3 Image Prompt Mistakes to Avoid

Even with a great template, three patterns consistently produce bad output. Spot them before you burn 50 generations.

Mistake 1: Vague Subjects

"A person" is not a subject — it is a category. The model has to invent the person from scratch, and the result is generic. A real subject has at least three identifying details: who they are, what they are wearing, and one piece of context. "A tired barista in a green apron, holding a steaming espresso." That is a subject. "A woman in a coffee shop." That is not.

Mistake 2: Skipping Lighting

Lighting is the single biggest factor in mood, and it is the easiest one to skip. A prompt that says nothing about lighting gets the model's default — flat, even, "product photography" lighting. If you want a mood (moody, dramatic, hopeful, warm), you have to describe the light. "Soft window light from camera left" is better than "moody." Specific lighting cues produce specific results.

Mistake 3: Trying to Get Text in the First Generation

Text rendering is the worst-supported feature in image generation. Midjourney is best at short words in stylized fonts. DALL-E handles short quotes in quotation marks. Flux is best with brand colors and hex codes. None of them are reliable for body copy. If you need a poster with a paragraph of text, the right workflow is to render the image first, then composite the text in a design tool. Trying to get 15 words right in a single generation is the fastest path to 50 wasted credits.

Putting It All Together

The 6-part formula — subject, action, environment, lighting, style, composition — is the spine. Model-specific tuning is the muscle. Midjourney wants short, comma-separated modifiers with parameters. DALL-E wants natural-language scene descriptions. Flux wants natural language plus specific camera and lens vocabulary. Pick the model by the use case, pick the format that fits the model, and run the 6-part checklist before you generate. One formula, three models, no magic keywords.

Try these techniques in PromptLab. Use the prompt builder to draft the 6-part formula, run it against multiple image generation models, and save the winning prompt as a reusable template.

Ask ChatGPT or Claude to "write a cold email" and you get a wall of polite, generic, instantly-skippable text. "I hope this email finds you well." "I wanted to reach out regarding." "Let me know if you have any questions." Every one of those lines trains the recipient to delete the message. The reason is not the model — it is the prompt. A vague prompt hands the AI every creative decision, and the AI defaults to the same corporate template it learned from millions of training emails. A structured prompt hands the AI only the decisions that matter — who is this person, what do they care about, what is the specific ask — and the result is an email that sounds like a human wrote it for that one specific reader.

This guide is the structured version. Five copy-paste email prompts for the five emails that eat the most time: cold outreach, follow-ups, internal updates, sales proposals, and customer support. Each template is built on the 5-part prompt structure from our general prompt engineering guide, tested across ChatGPT, Claude, and Gemini, and tuned with the model-specific tweaks from our Claude vs GPT-5 vs Gemini comparison. For the marketing-angled version of these prompts — landing pages, ad copy, social — see 10 Prompt Templates Every Marketer Needs.

Why "Write Me an Email" Is Not a Prompt

Email is one of the highest-context tasks you can throw at an LLM. The model has to infer the relationship between sender and recipient, the prior history, the desired outcome, the right tone, the right length, and the right level of formality — all from a one-line instruction. The default behavior is to assume "business professional, neutral, generic," which is exactly the voice that gets archived. The fix is the same as it is for any other prompt: name the output, define the constraints, give an example, and tell the model what success looks like.

Regie.ai's research on AI prompts for sales emails makes the point: AI prompts enable hyper-personalization at scale because the model can synthesize prospect data, persona, and intent into a single message that reads like a one-to-one email. The output quality is a function of input specificity. The "system prompt" sets the tone, voice, and rules for the whole sequence; the "user prompt" fills in the specific prospect, hook, and ask. Most "this AI email sounds robotic" complaints are actually "this prompt was too vague" complaints in disguise.

There is one more thing to internalize before the templates. Reply rate is not the same as open rate. An email can get a 70% open rate and a 0% reply rate if the ask is unclear, the next step is missing, or the subject promised something the body did not deliver. Every prompt below is built to optimize for reply, not open — which means a clear CTA, a low-friction next step, and a reason to respond that is not "just wanted to follow up."

The 5-Part Email Prompt Framework

Before the templates, here is the universal structure. Every email prompt should fill in all five parts. If you can only fill in three, the model will fill in the rest — and the default is the kind of email that gets archived.

1. Role & Recipient — who is writing, who is reading, and what is the relationship. "I am a SaaS founder writing to a VP of Sales at a 50-100 person company" is a role. "I am writing a sales email" is not.
2. Context & Hook — what just happened, or what specific thing about the recipient triggered this email. A trigger event (new funding, a job change, a published post, a recent purchase) is the single biggest lever on reply rate.
3. Value Proposition — what is in it for the recipient, in one sentence. Not a feature list, not a company bio. A specific outcome: save 4 hours a week, close 20% more deals, cut support response time in half.
4. Ask / CTA — the specific, low-friction next step. "Open to a 15-minute call next week" beats "let me know if you want to chat." A calendar link beats "find time on my calendar."
5. Format & Tone Constraints — length (under 80 words for cold email, 40-60 for follow-up), tone (casual, formal, peer-to-peer), signoff, and any specific phrases to avoid.

That is the framework. The next five sections are the templates — one for each of the five emails that eat the most time.

1. Cold Outreach Prompt

The hardest email to write is the first one. The recipient did not ask for it, does not know you, and has 200 other unread messages. The job of a cold email is not to close a deal — it is to earn a 6-line reply. Here is the prompt:

You are a B2B sales rep at a SaaS company. Write a cold outreach email under 80 words.

ROLE: I am writing to a [job title] at a [company size] company in [industry].
HOOK: They [specific trigger event — e.g., "just announced a Series B led by Accel" / "published a post on reducing support response time" / "hired their first RevOps lead last month"].
VALUE: We help companies like [reference customer] [specific outcome — e.g., "cut sales email drafting time by 70% while lifting reply rates 2x"]. One specific result: [reference customer] went from 4% to 11% reply rate in 6 weeks.
CTA: Open to a 15-minute call next week? I will send a specific agenda, not a generic demo.
TONE: Peer-to-peer, no fluff, no "I hope this finds you well." No exclamation marks.
SIGN-OFF: [Your name], [one-line title]

Example Output

Subject: 11% reply rate in 6 weeks

Hi Sarah,

Saw you just hired your first RevOps lead at Northwind — congrats. Most teams in that window spend 6-8 hours/week drafting outbound that converts at 3-4%.

We helped Replicant (similar profile, 60 reps) go from 4% to 11% reply rate in 6 weeks, while cutting drafting time by 70%. The short version: structured prompts + a sequenced follow-up framework, not magic.

Open to a 15-minute call next week? I will send a specific agenda — what to do in the first 30 days, what to skip.

Best,
Alex Chen
Head of Outbound, Acme

P.S. The Replicant teardown is here if you want the data first: [link]

2. Follow-Up Prompt (Bump with Value)

Most follow-ups fail because they are "just checking in." The recipient's inbox has 30 other "just checking in" messages. The follow-up that gets a reply is the one that adds a new piece of context, a new data point, or a new reason to respond. Here is the prompt:

You are a B2B sales rep following up on a cold email that received no reply. Write a follow-up under 60 words.

ORIGINAL EMAIL: [paste the first email you sent]
DAYS SINCE: [3-5 days is the proven range]
NEW CONTEXT: [pick ONE: a relevant case study, a new trigger event, a short relevant insight, a question that reframes the original ask]
TONE: Even shorter and lower-friction than the first email. Assume the recipient is busy and skeptical. No "just wanted to bump this up" or "circling back."
CTA: Make the ask smaller. "Worth a 10-minute look?" beats "open to a 15-minute call?" A specific yes/no question beats an open-ended one.

Example Output

Subject: Re: 11% reply rate in 6 weeks

Hi Sarah,

One more data point since my last note — we just published the Replicant teardown with the exact prompt templates and sequencing. Took them from 4% to 11% in 6 weeks.

Worth a 10-minute look? If it's not a fit, I'll stop the sequence.

Best,
Alex

The pattern: new context (the teardown), smaller ask (10 minutes vs. 15), explicit out (stop the sequence). That last line is the most underrated move in cold email — giving the recipient permission to say no. It paradoxically raises reply rates because it signals confidence and reduces the perceived cost of responding.

3. Internal Update Prompt (Status Report / Weekly Recap)

Internal status emails are the most-written, least-read email in most companies. The reason: they bury the lead in 5 paragraphs of context that the writer needs but the reader does not. The fix is the "BLUF" pattern — Bottom Line Up Front, then supporting context. The prompt:

You are a team lead writing a weekly status email to your skip-level and 3 peers. Write under 200 words.

BLUF (1-2 sentences): The single most important thing that happened this week. State it as a result, not an activity. "Shipped the v2 onboarding flow, lifted day-1 activation from 38% to 52%" beats "Worked on onboarding."
WINS (3 bullets): The 2-3 things that went well, each with a metric or concrete outcome.
BLOCKERS (1-2 bullets): What is stuck and what you need to unstick it. Name the specific person or decision.
NEXT WEEK (2-3 bullets): The 2-3 priorities for next week, in priority order.
TONE: Direct, no hedging, no "I think we made good progress." Skip-levels and peers scan, not read.

Example Output

Weekly Update — Growth Team

BLUF: Shipped the v2 onboarding flow Tuesday. Day-1 activation lifted from 38% to 52% in 4 days, with no change to week-1 retention yet (data still early).

Wins:
- Activation experiment reached 95% statistical significance at n=4,200
- Closed 2 design hires (Riley, Sam) — start July 1
- Hit 91% of Q2 OKR commitment on new-user activation

Blockers:
- Need pricing-page redesign final review from Brand before Friday launch
- Engineering capacity on referral feature is 50% of plan — proposing we deprioritize to Q3

Next week:
- Launch pricing-page redesign (Friday)
- Ship referral program v1 if eng capacity holds, otherwise descope
- Kick off Q3 planning (offsite Mon-Tue)

Reply by EOD Friday if you want to change any of the above.

The structure: lead with the result, group by category, end with a clear reply deadline. The "reply by EOD Friday" line is the highest-leverage addition — it converts a status email into a decision-making artifact, which is what leadership actually wants.

4. Sales Proposal Email Prompt

Sales proposal emails fail for a predictable reason: they summarize the proposal instead of selling the decision. The recipient does not need a recap of the deck — they need a reason to act this week instead of next quarter. The prompt:

You are an account executive sending a post-demo proposal email. Write under 250 words.

RECIPIENT: [decision-maker role and title — the person who signs, not the user]
WHAT WE AGREED IN THE DEMO: [2-3 specific points — their pain, the use case they want to solve first, the success metric they named]
PROPOSAL SUMMARY: [3 bullet points — what they get in phase 1, what it costs, what the outcome looks like in 90 days]
RISK OF INACTION: [one sentence — what happens if they wait another quarter]
CTA: Specific next step with a specific date. "Sign by Friday to start the kickoff the following Monday" beats "Let me know what you think."
TONE: Confident, direct, no "I hope this proposal meets your needs." Treat the proposal as a done deal waiting for a signature, not a request for approval.

5. Customer Support Reply Prompt

Customer support emails fail for the opposite reason: they are too apologetic and not informative enough. The reader has a problem, has already waited 4 hours, and is scanning for an answer — not for "I sincerely apologize for the inconvenience." The prompt:

You are a customer support agent replying to a customer who is frustrated. Write under 180 words.

CUSTOMER ISSUE: [what is broken, in their words, not yours]
ROOT CAUSE: [what actually happened — bug, misconfiguration, policy gap, misunderstanding]
WHAT WE ARE DOING: [specific action, with a specific timeline — "fixing in the v2.3 release on July 5" beats "we are looking into it"]
WHAT THEY CAN DO NOW: [workaround, if any, or "nothing needed on your end, we have it"]
TONE: Warm, direct, no "I completely understand your frustration." Acknowledge the problem in one sentence, then move to the fix. Do not over-apologize — over-apologizing trains customers to expect a refund.

Example Output

Hi Marcus,

You are right — the export failed on rows 14,022 through 18,500. That is a known bug in the v2.2 release that we patched yesterday.

What we are doing: the fix is live in v2.3, deploying to your account automatically at 6pm PT today (in 4 hours). Once it deploys, re-run the export — it will complete all 18,500 rows.

What you can do now: nothing. Your data is intact on our side; the export is the only thing that failed.

I will send a confirmation the moment the fix deploys. If you want a direct line to me for the next 24 hours, my phone is below.

Best,
Priya
Customer Engineering

The structure: name the issue, name the fix, name the timeline, name the workaround (or "nothing needed"). The "direct line to me" addition is the move that turns a 1-star ticket into a 5-star review. The customer does not actually want to call you — they want to know they could if they needed to. FindSkill's research on AI email templates for customer service agrees: the emails that resolve tickets fastest are the ones that follow a six-part structure (situation, root cause, action, timeline, workaround, signoff) and skip the apology theater.

3 Email Prompt Mistakes to Avoid

Even with great templates, three patterns consistently produce bad emails. Catch them before you send.

Mistake 1: "I Hope This Email Finds You Well"

This phrase is the most-deleted opening line in email history. The recipient knows you do not hope that. The recipient knows you copy-pasted the line. The recipient has read it 200 times this week. Open with the specific reason for the email, not with a wish for their wellbeing. "Saw your post on reducing support response time" is a better opening. "Saw your post" is even better — the recipient's eyes will go to the next line to find out which post.

Mistake 2: Burying the Ask

The biggest reply-rate killer is the email that ends with "let me know if you have any questions" instead of a specific ask. "Any questions" is a question the recipient can answer with no action. A specific ask — "open to a 15-minute call next Wednesday at 2pm PT?" — requires a yes or a no. The easier the response, the more likely it is to happen. PhantomBuster's research on ChatGPT prompts for sales emails makes the same point: the most effective prompts use personalization tokens specifically to make the CTA specific to the recipient.

Mistake 3: Forgetting the P.S.

The postscript is the second-most-read line in an email (after the subject line). Use it. A P.S. with a specific proof point, a case study link, or a low-friction reframe of the ask will outperform the same content in the body. "P.S. The Replicant teardown is here if you want the data first" is a way to add value for the recipient who is interested but not ready to commit. The P.S. is also the right place for the "if not, tell me to stop" line that paradoxically raises reply rates.

Putting It All Together

The 5-part email prompt framework — role & recipient, context & hook, value proposition, ask, format & tone — is the spine. The five templates above are the muscle. The biggest reply-rate lever in any of them is specificity: the recipient's name, the trigger event, the specific outcome, the specific ask. Pick the template that matches the email, fill in the five parts with the recipient in mind, and let the AI handle the prose. One framework, five emails, no more "I hope this finds you well."

Try these techniques in PromptLab. Use the prompt builder to draft a 5-part email prompt, run it against multiple LLMs side-by-side, and save the winning version as a reusable template for your next sequence.

Open the GPT Builder, type "make me a marketing assistant," and watch the most expensive part of Custom GPT creation go wrong in four seconds. The model happily writes a paragraph of friendly, vague, completely useless instructions that will haunt your GPT for the rest of its life. The reason is not the builder's quality — the builder is competent. The reason is that how to create a system prompt is a craft, not a fill-in-the-blank form, and the default builder output is the LLM equivalent of corporate lorem ipsum. A great system prompt has four distinct jobs to do, and a fifth one most builders forget. Get those four parts right and the same model that gave you "I am a helpful assistant that can help with marketing tasks" gives you a focused, on-voice, scope-limited, format-consistent tool that produces useful output on the first try and every retry after.

This guide is the full playbook. It is built on the 4-part anatomy of a production-grade system prompt — persona, scope, output contract, and failure modes — with the model-specific adjustments that make it work on ChatGPT's GPT-4o, GPT-4.1, and GPT-5. The structure is the same as the 5-part framework in our general prompt engineering guide, but applied to the long-lived "instruction" field that lives above every conversation, not the short-lived "user message" of a single turn. For the model-specific differences between ChatGPT, Claude, and Gemini, see Claude vs GPT-5 vs Gemini prompting; for the Claude-specific version of this anatomy, see 9 Tips to Write a Claude Prompt That Actually Works.

What a System Prompt Actually Is (and Is Not)

A system prompt is the invisible block of instructions the model reads before it reads the first user message. In a Custom GPT, it lives in the "Instructions" field of the GPT Builder. In the API, it goes in the system message of the messages array. Across all three, the job is identical: tell the model who it is, what it is allowed to do, what it is not allowed to do, and what shape the output should take — once, at the top of every conversation, for the entire life of the GPT.

That last part is the one most people miss. A system prompt is not a one-shot prompt — it is a long-lived behavior contract. Every conversation your user has with your Custom GPT, the system prompt is silently shaping the response. A vague system prompt does not just produce one bad answer — it produces thousands, multiplied by the number of users who will ever open that GPT. OpenAI's GPT builder guidance treats the instructions field as the single highest-leverage piece of the build, and the OpenAI Developer Community's "Custom GPT Limits" thread documents that the typical failure mode is not a weak model — it is a system prompt that delegates too many decisions to the model at runtime.

The 4-Part Anatomy of a Great System Prompt

Every system prompt that ships to production follows the same four-part structure. If you skip any of them, the model fills in the gap with its training-data default — which is generic, corporate, and on nobody's brand voice.

1. Persona & Voice — who the model is, what it knows, what tone it uses, and what it is forbidden from sounding like. The single biggest mistake is the "helpful assistant" persona, which is the model's default and adds nothing.
2. Scope & Capabilities — what the model is allowed to do, what it must always do (e.g., cite sources, ask for clarification, refuse off-topic requests), and what it must never do regardless of how the user phrases it.
3. Output Contract — the exact shape of the response: format (markdown vs. plain), length (1 paragraph vs. 5 bullets vs. JSON), required sections, and what to do when the model does not know the answer.
4. Failure Modes & Anti-Patterns — the specific bad behaviors the model tends to fall into on this task, and the explicit instruction to avoid them. This is the section most builders skip and the one with the highest quality lift.

1. Persona & Voice: "Who Is This GPT, Exactly?"

The persona section answers "if this GPT were a person, who would they be?" The wrong answer is "a helpful AI assistant." The right answer is specific enough that you could draw a face: their expertise, their personality, their default tone, and the persona they are explicitly not. The OpenAI Developer Community thread on "Custom GPT Instructions" finds that 2nd-person ("You are…") and 3rd-person ("This GPT is…") both work, but 2nd-person produces slightly more consistent persona adherence because it reads as direct onboarding language to the model.

PERSONA:
You are a senior B2B copywriter with 12 years of experience writing
for early-stage SaaS companies. You have shipped copy for Notion,
Linear, and Figma in their first 100 customers phase. Your default
tone is direct, peer-to-peer, slightly opinionated, and allergic
to corporate jargon. You write the way a sharp founder would write
to another sharp founder.

YOU ARE NOT:
- A general "helpful assistant"
- A marketing agency voice (no "we leverage synergies")
- An academic voice (no "it is important to note that")
- A hype voice (no "this is a game-changer")

The "YOU ARE NOT" block is doing as much work as the "PERSONA" block. Without it, the model drifts toward its training-data average for "B2B copywriter" — which is the agency voice. The negative constraints are what keep the model on the persona you actually want.

2. Scope & Capabilities: What This GPT Does and Does Not Do

Scope is where most Custom GPTs fail silently. The model has no way to know what your GPT is "for" unless you say so, and no way to know what to refuse unless you tell it. A scope-less GPT will cheerfully help with anything, which means it will be useless at the one thing it was built for. The "Custom GPT Limits" research on OpenAI's community forum documents that the most common production failure is scope creep — the GPT drifts off-topic into a generic chatbot because the system prompt did not bound the domain.

IN SCOPE — you will help with:
- Landing page copy (hero, subhead, 3 benefit bullets, CTA)
- Cold email sequences (3-email cadence, 80 words per email)
- Pricing page copy (plan names, value props, FAQ entries)
- Sales one-pagers (problem, solution, proof, ask)

ALWAYS DO (no matter what the user asks):
- Ask 1 clarifying question if the request is vague
- Cite the specific input you used ("Based on the trigger
  event you gave, I opened with…")
- Produce 2-3 variants, not 1, for landing page and email

REFUSE / REDIRECT (do not do these even if the user asks):
- Do not write for B2C, ecommerce, or consumer brands
- Do not write in any language other than English
- Do not generate images, code, or anything outside copy
- Do not make up customer logos, statistics, or testimonials

The "ALWAYS DO" block is the part most builders skip. It is the difference between a GPT that follows your workflow and a GPT that invents its own. If your workflow requires the model to ask a clarifying question before writing, say so explicitly — the model will not infer it from the persona alone.

3. Output Contract: The Shape of Every Response

The output contract is the most underrated section. Most system prompts leave the response shape up to the model, and the model defaults to "a friendly paragraph of prose." For a copy GPT, that is the wrong format — the user wants variants, not a paragraph. For a code-review GPT, the user wants a diff. For a research GPT, the user wants a structured summary with citations.

OUTPUT FORMAT:
- Use markdown headings (##) for each deliverable
- Use bullet points for variants, not numbered lists
- Put the recommended option FIRST, then 2 alternatives
- End every response with "Want me to adjust the [tone |
  length | audience]?" — one line, never a paragraph

LENGTH:
- Landing page: 1 hero, 1 subhead, 3 bullets, 1 CTA, < 80 words
- Cold email: 3 emails, < 80 words each
- Pricing page: 3 plans, < 30 words per plan
- One-pager: 4 sections, < 200 words total

WHEN YOU DON'T KNOW:
- If the user has not given you enough input, ask 1
  clarifying question. Do not guess.
- If the user asks you to write for a brand you do not
  have information about, ask for the brand's existing
  copy as reference.
- Never invent statistics, customer logos, or product
  features.

The "WHEN YOU DON'T KNOW" sub-block is the single highest-quality section in the whole system prompt. Models hallucinate by default. Telling the model explicitly "never invent statistics" cuts the hallucination rate on that specific failure mode by a large margin in practice. The same principle applies to any other category the model is likely to make up: code, citations, customer names, dollar figures, dates.

4. Failure Modes & Anti-Patterns: Catch the Bad Habits Before They Ship

Every domain has its own failure patterns. For B2B copy, the model tends to: (1) start with "I hope this email finds you well," (2) bury the ask in the third paragraph, (3) invent product features, (4) end every email with "let me know if you have any questions." For a research GPT, the failure patterns are: citing papers that do not exist, summarizing the abstract instead of synthesizing, presenting the model's opinion as fact. For a tutoring GPT: giving the answer instead of guiding, using language above the student's level, skipping the worked example. Name them in the system prompt.

DO NOT:
- Open with "I hope this email finds you well" or any other
  empty pleasantries
- Use "in today's fast-paced world", "in the ever-evolving
  landscape", or "in the realm of"
- Use the word "leverage" as a verb
- End with "let me know if you have any questions"
- Use exclamation marks (they read as junior)
- Bury the CTA in the third paragraph — CTAs go in the
  last line, always
- Produce a single variant when the user asked for copy

CustomGPT.ai's research on production Custom GPTs finds that the anti-pattern block is the single biggest quality lift in a system prompt, because the model has seen every one of those bad patterns thousands of times in training data and needs an explicit instruction to override them. The patterns above are real, common, and exactly the kind of thing a copywriter would refuse to ship. The system prompt is the only place you can stop them at the source.

Putting It All Together: A Production-Ready System Prompt

Here is the complete system prompt assembled from the four parts above. It is around 250 words, which is the right size for a Custom GPT in this category — long enough to be specific, short enough that the model actually follows every line. Past 600 words, the model starts ignoring parts of the instructions; past 1,000, the effect is severe. If you find yourself past 600 words, you probably have a workflow problem (a chain of GPTs, or a system prompt plus a knowledge file) and not a system prompt problem.

PERSONA:
You are a senior B2B copywriter with 12 years of experience writing
for early-stage SaaS companies. You have shipped copy for Notion,
Linear, and Figma in their first 100 customers phase. Your default
tone is direct, peer-to-peer, slightly opinionated, and allergic
to corporate jargon.

YOU ARE NOT: a helpful assistant, a marketing agency, an academic,
or a hype voice.

IN SCOPE: landing page copy, cold email sequences, pricing page
copy, sales one-pagers for B2B SaaS only.

ALWAYS DO: ask 1 clarifying question if the request is vague; cite
the specific input you used; produce 2-3 variants for landing pages
and email.

REFUSE: B2C, ecommerce, non-English, code, image generation, or
any request to invent logos/statistics/testimonials.

OUTPUT FORMAT: markdown, bullets for variants, recommended option
first, end with "Want me to adjust the [tone | length | audience]?"

LENGTH: landing page < 80 words, cold email 3x80 words, pricing
page 3x30 words, one-pager < 200 words.

DO NOT: open with pleasantries, use "leverage" as a verb, end with
"let me know if you have any questions", use exclamation marks,
bury the CTA, or produce a single variant when copy is requested.

Drop that into the GPT Builder Instructions field, test it against 5 real user inputs (not "write me a landing page" but a real one with a real product, a real buyer, and a real trigger event), and refine the parts that miss. Two common refinements: if the model is too verbose, tighten the LENGTH block. If the model keeps producing off-brand tone, add another "YOU ARE NOT" line. The system prompt is a living document — version it, change one section at a time, and A/B test the changes with real inputs.