AI CEO LabHow To Write Better AI Prompts: Prompt Like An Operator
Every prompt on this site has the same five parts: a role, the context pasted in, the constraints it may not cross, the exact shape of the output, and, when it matters, an interview instead of a guess. Here is the pattern, why each part exists, and the one that most people leave out.
Most people prompt the way they text. A sentence, a hope, a reroll. Then they decide the model is not that good. The model is fine. It was asked for the average version of a vague idea and it delivered exactly that.
I learned the clearest version of this from a video course I took, where the lesson was that a prompt is direction, not description. The less you direct, the more the model fills in, and when it fills in it picks the most ordinary choice every time. That course taught a six part frame for images. I run a five part frame for business, and it is in every prompt on this site.
Look at any of them. Act as my landing page strategist. Here is the ad, the offer, the audience. Do not invent any link or claim. Produce, numbered, these seven things. Interview me one question at a time. Role, context, constraints, output shape, interview. Same bones, every time, whether it is an ad audit, a DM bot or a newsletter issue.
This guide is the pattern with the reason behind each part, the part almost everyone skips, and the difference between a prompt you paste once and an instruction that stands. The file that carries your judgment into every prompt is the next guide, Turn Your Experience Into One AI File.
The three levels
One sentence, a reroll, a second sentence, a copy of whatever came out third.
Every prompt carries a role, the real context pasted in, the lines it may not cross, a numbered output shape, and an interview when the model would otherwise guess. Results come out usable the first time.
The role and constraints live in saved instructions, the context comes from a file the agent reads, and your prompts shrink to the trigger phrase and the new information. The interview runs itself and stops for your yes before anything live.
The mental model
Describe what you want in a sentence and see what comes back.
Direct. Give it a job, the real material, the lines it may not cross, the shape of the answer, and make it ask before it guesses.
| Role | Talks to you | Job |
|---|---|---|
| Role | First line | Act as my X. Sets the job and the standard, not a personality. |
| Context | Pasted, not summarized | The actual ad, offer, notes, numbers. Never a description of them. |
| Constraints | Before the ask | Do not invent. Read only. Exact yes before anything live. No em dashes. |
| Output shape | Numbered | Produce these N things in this order. A table. One paragraph. Never leave it open. |
| Interview | When it would guess | One question at a time. Refuse to continue until specific. |
The five parts, in order
1. Role: give it a job, not a personality
Act as my ad account auditor. Act as my GoHighLevel automation builder. Act as my newsletter editor. The role sets the standard the answer is judged against. It is not you are a world class genius, which changes nothing. It is the job title of the person you would hire for this, in your business, today.
2. Context: paste it, never describe it
The single biggest difference between my prompts and the ones I see. Here is the ad I am running, pasted. Here is the offer with the price and what is included. Here are my notes from the week. A description of your material makes the model guess what the material says. The material itself removes the guess. If it is too long to paste, that is what a file is for.
3. Constraints: the lines it may not cross
Do not invent any link, number or claim. Read only, never pause, edit or spend. Never change a budget without my exact yes. Use only numbers I gave you. No em dashes. Every one of these lines exists because the model once did the thing. Constraints go before the ask, and the important ones get repeated at the end, because the end of a prompt is what the model reads last.
4. Output shape: never leave it open
Produce, numbered: one, the tag list; two, the workflow as trigger and steps; three, the JSON payload; four, the test. Or a table with these columns. Or one paragraph. An open ended prompt gets an essay. A shaped prompt gets something you can paste into the tool. The shape is where most of the time saving lives.
5. Interview: make it ask before it guesses
The part almost everyone skips. When the answer depends on things only you know, the prompt says interview me one question at a time, push back on vague answers, refuse to continue until each is specific. The judgment file prompt is nothing but this. The stack test refuses to continue until four things are written. An interview turns a guess into your answer, and it takes two minutes.
6. Say it out loud, then cut
I dictate most prompts. Speaking forces plain words and one person on the other side. Then I cut anything that is not one of the five parts. The prompts on this site are short on purpose, because the long ones drift. If a prompt has a paragraph of motivation, delete the paragraph.
Starter prompts
Paste these as written. They are short on purpose, because the long ones drift.
Act as my prompt editor. Here is a rough ask I want an AI to do: [paste]. Rewrite it in five parts and label each: (1) Role, as a job title in my business, one line; (2) Context, listing exactly what I must paste in, not describe, for the answer to be real; (3) Constraints, the lines it may not cross, including do not invent any link, number or claim, and an exact yes before anything live if the task touches a live system; (4) Output shape, numbered or a table, in the order I will use it; (5) Interview, the questions it must ask me one at a time before answering, refusing vague answers. Then list every place my original ask would have forced a guess. Keep it short. No em dashes.
Before you answer, interview me one question at a time. Push back on any answer that is generic the way you would push a new hire who says they just know it when they see it. Do not continue to the next question until the current answer contains a specific number, name, date or example. When every question is answered, produce the output in the shape above.
From now on in this project: act as my [role]. Never invent a link, number, result or claim; if you do not have it, say so and ask. Treat every connected account as read only. Any send, publish, spend or record change waits for my exact yes in a separate message. Use only the numbers I give you. Plain words, short paragraphs, no em dashes. End every answer with the one thing that needs me.
The pasted prompt, then the standing instruction
Layer 1 is the five part prompt pasted into a chat window. It works anywhere and it is how every prompt on this site is written. Layer 2 is moving the parts that never change out of the message and into a place the agent reads every time, so the prompt you type becomes the trigger and the new information. These are the settings.
Role and constraints go in the instructions
A Claude Project's instructions, a custom GPT's instructions, a skill file, or the config file an agent reads on start. Once there, they apply to every message without being retyped.
Context comes from a file
The company operating file, the stack file, the offer page. Attached to the Project or in a folder the agent can read. Paste only what is new this time.
Output shape lives in the skill
A skill file ends with the report format. Every run returns the same shape, which is what makes week over week comparison possible.
The interview is a trigger phrase
The skill's description line lists the phrases that start it. You say the phrase, it asks the questions.
Live actions stay gated by permission, not by prompt
A constraint that says read only is good. A token that can only read is better. Do both. The agent safety guide covers the gate.
Give it to your agent, three ways
Same skill, three worlds. Pick the one you actually use. The skill file at the bottom of this page is the instructions in every case.
An agent with connectors (Claude Desktop, Claude Code, Grok Bot)
No connector is needed to prompt well. What changes with Claude is where the parts live: Projects hold instructions and files; Claude Code reads a config file on start and loads skills by trigger phrase.
- Claude Projects: paste the role and constraints as the project instructions, attach the context files, and type only the ask and the new material.
- Claude Code: put the role and constraints in the project's config file, save each recurring prompt as a skill file with the output shape at the end, and say the trigger phrase.
- Paste the skill at the bottom of this page and say shape this prompt to rewrite any rough ask into the five parts.
- For anything that acts on a live system, keep the exact yes line in the prompt and the read only scope in the connector's permissions.
# a standing instruction file an agent reads on start cat > CLAUDE.md <<'EOF' Act as my operator. Constraints: never invent a link, number or claim; read only on every connected account; any send, publish, spend or record change waits for my exact yes; no em dashes. Context: read stack.md and company-os.md before answering business questions. Output: numbered, in the order asked, and a one line summary at the end. EOF
ChatGPT (a project or a custom GPT)
The same shape works in ChatGPT, with a Project or a custom GPT carrying the constant parts.
- Create a Project. Paste role and constraints as the instructions. Upload the context files.
- Type only the ask, the pasted new material and the output shape.
- For a recurring prompt, make a custom GPT whose instructions end with the output shape and whose conversation starters are the trigger phrases.
- Keep the exact yes line in any prompt that would lead to a live action, because ChatGPT Actions can write if you let them.
Anything with an API (a token and a curl call)
If you call a model from code, the five parts map to the request: role and constraints in the system prompt, context and the ask in the user message, output shape as a schema or a format instruction. The interview becomes a loop that asks until the required fields are filled.
- System prompt: the role and every constraint, verbatim from your standing instructions file.
- User message: the pasted material and the ask, with the output shape at the end.
- Ask for the output as JSON with named fields when a tool will read it, so the shape is enforced instead of hoped for.
- For the interview, keep a list of required fields and re-prompt for each empty one before the final call.
export CLAUDE_MODEL=... # the current model id from the Anthropic docs
curl https://api.anthropic.com/v1/messages \
-H "x-api-key: $ANTHROPIC_API_KEY" -H "anthropic-version: 2023-06-01" -H "content-type: application/json" \
-d '{"model":"'"$CLAUDE_MODEL"'","max_tokens":1200,
"system":"Act as my ad account auditor. Read only. Never invent a number. If a value is missing, ask for it.",
"messages":[{"role":"user","content":"Here is last week at ad level: [paste]. Produce JSON with fields score, fixes (3, ranked by dollars), winner, kill_list."}]}'Failure modes
Every one of these has happened to me or to someone I set this up for.
| Failure | Fix |
|---|---|
| Generic answer that could be for any business | Paste the material. A description of it is a guess about it |
| An invented link, number or client name in the output | The constraint was missing or buried. Put it before the ask and repeat it at the end |
| An essay when you needed a list | Number the output. Say the order |
| The model guessed your price, your audience, your rule | Add the interview line. One question at a time, refuse vague answers |
| Retyping the same three constraints every message | Move them to Project instructions or a skill file |
| A prompt with a paragraph of motivation and no shape | Delete the paragraph. Keep the five parts |
| An agent sent something because the prompt said to draft it | Exact yes line in the prompt, and a scope that cannot send, in the permissions |
The tools I use for this
| Tool | What it is for here | |
|---|---|---|
| Claude | Where I prompt, and where Projects and skill files hold the parts that never change. | Open |
| Wispr Flow | I dictate prompts. Speaking forces plain words and one person on the other side. | Get it |
| ChatGPT | Same shape, Projects and custom GPTs for the constant parts. | no link, just use it |
The free skill
It rewrites any prompt into the five part shape. Give it a rough ask and it returns the role, the context it needs from you, the constraints it should carry, a numbered output shape, and the interview questions it should ask before answering, then flags every place the original prompt would have forced a guess.
--- name: prompt-like-an-operator description: Rewrites any rough ask into the five part operator prompt: role as a job title, context to paste rather than describe, constraints it may not cross, a numbered output shape, and an interview that runs before the model guesses. Flags every place the original would have forced a guess. Trigger on "shape this prompt", "fix my prompt", "why is the answer generic", "make this a standing instruction". --- # Prompt The Way An Operator Prompts A prompt is direction, not description. The less direction it carries, the more the model fills in, and it fills in the most average version every time. Every prompt you produce has five labeled parts, in this order. ## The five parts ### 1. Role One line: "Act as my [job title]." The job the owner would hire for this, in their business, today. Sets the standard the answer is judged against. Never a personality, never "world class", never "genius". ### 2. Context List exactly what the owner must paste in, not describe, for the answer to be real: the ad, the offer with price and inclusions, the notes, the numbers, the page. If it is too long to paste, it belongs in a file attached to the project or read by the agent. A description of the material makes the model guess what the material says. ### 3. Constraints The lines it may not cross, placed before the ask and repeated at the end. Always include: - Do not invent any link, number, result, client name or claim. If missing, say so. - Use only the numbers the owner gave. - No em dashes. And when the task touches a live system: - Treat every connected account as read only. - Any send, publish, spend or record change waits for an exact yes in a separate message. ### 4. Output shape Numbered items in the order the owner will use them, or a table with named columns, or one paragraph. Never open ended. When a tool will read the output, ask for JSON with named fields so the shape is enforced instead of hoped for. ### 5. Interview When the answer depends on things only the owner knows, add: "Interview me one question at a time. Push back on generic answers. Do not continue until the current answer contains a specific number, name, date or example." List the questions. The judgment file interview and the stack test are examples of prompts that are nothing but this. ## What you return 1. The rewritten prompt with the five parts labeled. 2. A list of every place the original ask would have forced a guess, one line each. 3. If the owner asked for a standing instruction: the role, constraints and output shape separated out as text for a Project's instructions, a custom GPT, a skill file or a config file the agent reads on start, with the note that the typed prompt then becomes the trigger phrase plus the new material only. ## Rules for the rewrite - Short on purpose. Delete any paragraph of motivation or praise. The long ones drift. - Plain words, written to one person. - Keep the owner's specific words where they appear. - A constraint in words is good; a permission that cannot write is better. Say so when the task touches anything live.
What done looks like at thirty days
- Your three most used prompts carry all five parts
- Role and constraints live in saved instructions, not in every message
- At least one context file is attached instead of pasted
- Every prompt that could touch a live system carries an exact yes line
- One prompt uses an interview and the answer changed because of it
- You have not rerolled a first answer in a week
Want your prompts turned into standing instructions for your own agent?
Inside the AI CEO Lab we build the file your agent reads on start and the skills that carry your constraints, so the prompt you type becomes one line.