What is the C.A.R.E. method for briefing AI?

C.A.R.E. is a 4-box canvas to structure a prompt for a generative AI (ChatGPT, Claude, Gemini, Mistral, etc.). C — Context (who you are, who you're addressing, what situation). A — Action (the precise task, with a clear verb). R — Rules (expected format, length, tone, what to avoid). E — Example (one or two models of what you consider a good result). The Example is the most neglected ingredient and the one that moves the needle most.

Why does 90% of an AI result's quality depend on the prompt?

Because a generative AI doesn't guess your context. It produces a statistically plausible answer from what you give it. Ask 'write a LinkedIn post' and it produces the average LinkedIn post. Tell it who you are, who you're speaking to, in what tone, and with an example of what works for you, and it produces something close to you. The prompt is the only lever you have.

How long does it take to write a C.A.R.E. brief?

Five to ten minutes for a reusable high-quality brief. Longer than 'write an email to my client', but this brief will serve you dozens of times. Store your best C.A.R.E. briefs in a library (Notion, Airtable, plain text file) — it's your most durable asset because it doesn't depend on any single tool. It works on ChatGPT today, on Claude's successor tomorrow.

How do I write a good Example in the C.A.R.E. canvas?

Give a real, complete, marked-up case. Template: 'Example of what I consider good: [paste 5-15 lines of your best past work]. What makes it good: [2-3 concrete criteria].' If you have no example handy, describe what makes a good result (length, structure, tone, things you never see in good ones). Showing beats telling, but telling beats nothing.

Does C.A.R.E. work the same on ChatGPT, Claude, Gemini and Mistral?

Yes. C.A.R.E. isn't model-specific — it's a brief structure. Every conversational model (ChatGPT, Claude, Gemini, Mistral Le Chat, Llama, etc.) responds better to a structured brief than to a vague prompt. Model-to-model differences are about tone, creativity, factual accuracy — not how to brief them. The same C.A.R.E. produces varied results across models, which is healthy: pick the model whose output is closest to what you want.

What are the most common traps when briefing AI?

Four recurring traps. (1) Skip Context: 'write a professional email' yields a generic email. (2) Action too vague: 'help me with my LinkedIn' says nothing — prefer 'write 3 LinkedIn post hooks'. (3) No Rules: length, tone, format missing — you get an indecisive output. (4) No Example: the costliest trap — you get the average output instead of yours. If one box is empty, so is the result.

C.A.R.E. method: brief ChatGPT, Claude or Gemini like a skilled intern

Q: Do you need a magic prompt or secret formulas to brief AI well?

No. The 'magic prompts' sold on social media are at best generic templates, at worst marketing. A good prompt is a brief, the way you'd brief a competent intern who knows nothing about your context. The C.A.R.E. canvas covers what's needed in four boxes. The rest is precision in each box — not a hidden formula.

TL;DR. A good AI prompt isn't a magic formula. It's a brief — the way you'd brief a competent intern who knows nothing about your context. The C.A.R.E. canvas structures that brief in four boxes: Context (who, for whom, what situation), Action (the task, a clear verb), Rules (format, length, tone, what to avoid), Example (one or two models of what works for you). The Example is the box everyone skips and the one that changes everything — showing beats telling. Works on ChatGPT, Claude, Gemini, Mistral and their successors.

Why your ChatGPT prompts deliver average results

You've been testing ChatGPT, Claude or Gemini for months. You sometimes get stunning outputs, but most of the time you get average outputs you have to rewrite.

The problem is almost never the model. It's the prompt.

A generative AI doesn't guess your context. It produces a statistically plausible answer from what you give it. Ask "write a LinkedIn post" and it produces the average LinkedIn post — the one a thousand people have already written. Tell it who you are, who you're speaking to, in what tone, and with an example of what works for you, and it produces something close to you.

The prompt is the only lever you have. And it's precisely the one nobody works on.

The C.A.R.E. canvas in four boxes

No formulas to memorize. Four boxes are enough:

C — Context

Who you are, who you're addressing, what situation. Brand/activity, audience, problem to solve, constraints.

Bad: "I'm a consultant." Good: "I'm an independent cybersecurity consultant in Nantes, France. I address non-technical SMB leaders (10-50 employees) who received an NIS2 questionnaire from their enterprise customer and don't know where to start."

A — Action

The precise task, with a clear verb. Not "help me" or "tell me about" — a verb and a concrete deliverable.

Bad: "Help me with my LinkedIn." Good: "Write 3 LinkedIn post hooks, 100-150 characters each, with at least one opening on a verifiable statistic."

R — Rules

Expected format, length, tone, what to avoid. This is where you frame the output.

Template: "Format: 3 numbered hooks. Length: 100-150 characters each. Tone: direct, plain-spoken, no emojis. Avoid: superlatives ('amazing', 'revolutionary'), corporate jargon, hashtags."

E — Example

One or two models of what you consider "good". The ingredient everyone neglects and the one that changes everything.

Template: "Example of a hook that worked for me: '60% of SMBs hit by ransomware shut down within 18 months. Here are the 3 measures that would have saved them.' What makes it good: opens on a verifiable statistic, concrete promise, dry tone without pathos."

The Example isn't a perfectionist's whim. It's what turns a generic output into yours.

A complete C.A.R.E. in action

Here's a real brief, ready to paste into ChatGPT, Claude, Gemini or Mistral:

Context. MAG&Cie is a Nantes-based consulting firm led by a freelance CTO/CPO since 2015. We publish on LinkedIn for SMB leaders and early-stage startup founders weighing internal vs outsourced tech leadership. Our angle: pragmatic, plain-spoken, anti-bullshit.

Action. Write 3 LinkedIn post hooks for an article arguing that a half-time outsourced CTO is more profitable than a full-time junior CTO for a 5-15 person startup.

Rules. Format: 3 numbered hooks, blank line between. Length: 120-180 characters each. Tone: direct, no emojis, no hashtags, no superlatives. At least one hook must open on a question, at least one on a statistic. No "revolutionary", "game changer", "amazing", or coachy language.

Example. Hook that worked recently: "Hiring a $80k junior CTO to do the work of a half-time senior is paying full price for half the result. Here's why." What makes it a good hook: states a concrete cost, flips common sense, clear promise of what's next.

Paste that into any conversational model: you get hooks close to your voice on the first try. That's the difference between 10 minutes of briefing and 2 hours of rewriting.

Store your C.A.R.E. briefs — your most durable asset

Your best briefs become your reusable templates. Keep them where you can find them (Notion, Airtable, plain text file) — not in ChatGPT history, which evaporates after three weeks.

It's your most durable asset. It depends on no tool. It works on ChatGPT today. It'll work on Claude's successor tomorrow. The day you switch your favorite model, your C.A.R.E. briefs move with you.

Build this library by category: Content, Prospecting, Customer Support, Admin, Monitoring, Misc. One sheet per recurring task is enough.

The 4 traps that ruin a C.A.R.E.

When a C.A.R.E. brief disappoints, it's almost always one of these four boxes that's empty or fuzzy.

Context missing or too short. "Write a professional email" yields a generic email. Specify who you are, who you're writing to, in what setting.
Action too vague. "Help me with my LinkedIn" says nothing to the AI. "Write X" or "Analyze Y and produce Z" are actionable.
Rules missing. Without format, length, tone, you get an indecisive output. A fuzzy rule beats no rule.
No Example. The costliest trap. Without an example, the AI produces the average result instead of yours. If you have no example handy, describe what makes a good output — weaker than showing, but better than nothing.

C.A.R.E. works the same on ChatGPT, Claude, Gemini, Mistral

C.A.R.E. isn't model-specific. It's a brief structure. Every consumer conversational model responds better to a structured brief than to a vague prompt.

Model-to-model differences play out elsewhere:

Tone and creativity. Claude tends to be more nuanced, ChatGPT more directive, Gemini more factual, Mistral more concise. Test the same C.A.R.E. on two models, keep the one whose output you like most.
Factual accuracy. For anything involving figures, dates, references, verify — whatever the model.
Confidentiality. Read the ToS. Don't paste sensitive client/HR/financial data into a model whose data handling you haven't verified.

Going further

The C.A.R.E. method has a natural companion: the Trigger → Action → Control breakdown to structure any AI automation, detailed in Automate your business with AI without coding: the solopreneur method.

To apply it to an assistant that triages your email, prepares your days and handles your meetings, read AI agent as a personal assistant: email triage, calendar, meetings.