How to write a prompt that actually works

Most CEOs write prompts like search queries. Short, decontextualized, and hopeful. There is a better way, and it takes less than a minute to learn.

What you'll have when you're done

A five-part briefing structure you can drop into any AI chat, three worked examples you can borrow right now, and a reusable template you save once and never have to think about again. The output from your prompts will look like it came from someone who already knows your business, not from a model that just met you.

The CEO typing search queries into a chat window

There is a pattern. You open Claude or ChatGPT, type something like "help me with my competitor research" or "write a board update," and get back a clean, confident, generic wall of text. You skim it. You close the tab. You decide the hype is overblown.

It is not overblown. The briefing is.

You are treating an LLM like a search engine. Search engines reward short queries because they are matching keywords to documents. Language models are doing something completely different. They are predicting the most useful response given everything you gave them. Give them nothing, and they have to guess. They will guess fluently, and they will guess wrong.

The context you provide matters more than the wording of your question. That is the durable insight behind every high-output prompt you will ever write. The anatomy of a prompt (task, context, format, constraints) is already defined; this article is about how to brief one so it actually works.

What you need first

• A Claude or ChatGPT account. If you are still setting up, your first 30 minutes with Claude covers the setup.
• One real piece of work you have been putting off or one you do weekly. That is your test case for everything in this article.
• Five minutes to read this once, and another five to run your first real prompt with the template below.

You do not need any technical background. Prompting is briefing, and you already know how to brief people.

Step by step

The method: brief the model like a sharp new hire

Think about the last time you brought on a capable new hire who came in knowing their craft but nothing about your business. You did not hand them a two-word task. You told them what the company does, who they are writing for, what decision the work feeds, and what you do not want. Same move here.

Five parts, in order. Put context before the task. Models weight the end of a prompt more heavily, so you want the task to land last against fresh framing you have just given them.

1. Role. Who the model is acting as. "You are a competitive-intelligence analyst who covers B2B SaaS" does more than "analyze this." It filters the vocabulary, the frame, and the level of precision.
2. Context. What the model cannot see: your industry, your company size, who reads the output, what decision this feeds. This is the part most people skip. It is the most important part.
3. Task. One specific deliverable. Not "help me with X," but "write a table," "draft a one-page memo," "list five competitors with these three data points."
4. Format. How the output should be structured: bullets vs. table vs. prose, word count, sections. Without this, the model picks a shape that may not fit what you need.
5. Constraints. What to avoid. "No hedging language." "Skip competitors with under $5M in funding." "No corporate speak." Constraints are the highest-leverage lines in any prompt.

Three worked examples

These are the center of this article. Read them slowly. The before versions are close to what most CEOs type. The after versions are what gets used.

Example A: Competitor research

Weak:

Tell me about my competitors.

Strong:

You are a competitive-intelligence analyst who covers B2B SaaS for mid-market companies. I run an $8M ARR project management tool for construction contractors. List the top 5 direct competitors. For each: their pricing tier aimed at contractors, one capability they have that I lack, and one weakness I could exploit. Use a table. Skip any competitor with under $5M in funding.

What changed: the role filters the analysis to the right level of sophistication. The context block tells the model what market it is in and who it is competing for. The deliverable is specific (a table, five rows, three columns). The constraint cuts noise before the model wastes tokens on it.

Example B: Board or investor update

Weak:

Draft a board update for me.

Strong:

You are a chief of staff who writes tight board memos. I need a one-page Q2 update for a five-person board of mostly financial investors. Context: revenue is $1.2M MRR (up 18% QoQ), we missed our hiring goal by 2 people, and we closed our first enterprise deal at $240K ARR. Format: three sections only, Wins, Misses, Next-quarter priorities. Bullets, not paragraphs. No hedging language. Max 250 words.

What changed: the role acts as a professional filter on the writing quality. The context block hands the model the facts it cannot guess. The format kills the "here is a comprehensive overview of your quarter" shape that no board member wants to read. "No hedging language" is the highest-ROI single line in this prompt. It kills the AI tell before it lands on your board's screens.

Example C: Hiring or job description

Weak:

Write a job description for a head of marketing.

Strong:

You are a senior recruiter who writes JDs that attract operator-minded, not brand-minded, candidates. I need a JD for a Head of Marketing at a bootstrapped DTC fitness supplement brand doing $4M/year. They will own paid acquisition, email, and influencer, reporting directly to me (the founder). Tone: candid, founder-led, no corporate speak. Format: three bullet sections, What you will own / What we are looking for / What we offer. Max 400 words.

What changed: the role filters out the HR-speak that makes JDs sound like every other JD. The business context calibrates the scope and seniority of the role. "No corporate speak" is the highest-leverage line. Without it, you get a JD that could belong to any company.

The reusable template

Copy this once. Save it somewhere you open daily. Paste it, fill in the brackets, and send.

You are [specific role or expertise].
I need [one concrete deliverable].
Context:
- Company / situation: [what the model cannot infer]
- Audience: [who reads or uses the output]
- Decision this feeds: [why it matters]
Format:
- Structure: [bullets / table / memo]
- Length: [word or section count]
- Tone: [candid / formal / founder-voice]
Constraints:
- [what to avoid]
- [assumptions to skip]

You do not need to memorize clever wording. Anthropic's and OpenAI's guidance, and MIT Sloan's research on how non-technical operators get the most from AI, all point to the same conclusion: a reusable briefing structure beats prompt-crafting tricks every time. The structure is the skill.

Three pro techniques worth knowing

Show the model what you want. Paste a short example of the output you are looking for. If you want bullets that sound like how you actually talk, paste two bullets you would have written. Giving the model a sample is the single highest-ROI addition you can make to any prompt, per Anthropic's and OpenAI's guidance on few-shot examples. One good example does more than three paragraphs of format description.

Make it ask you questions first. Add this line to any draft request: "Before you draft, list the three things you'd need to know to write this well." The model will surface the gaps in your brief, and you will be surprised how often one of them is something you actually know but forgot to include. This one move cuts revision cycles in half.

Tell it to think step by step for analysis and judgment calls. Anthropic's best-practice guidance recommends explicit step-by-step reasoning instructions for prompts that involve evaluating options, diagnosing situations, or making recommendations. "Walk through your reasoning step by step before you write the recommendation" works. Skip this instruction on formatting and drafting tasks where it just adds padding.

How you'll know it's working

The output uses your specifics. Not "here are some common considerations for Q2 board updates" but your actual MRR number, your actual miss, your actual next-quarter priority named by name. If the output could belong to any company, your context block is empty. Go back and fill it.

You use the output directly. You forward it, you paste it into a doc, you read it in the meeting. You do not spend fifteen minutes rewriting it into something useful. If you are rewriting heavily, one of two things happened: your constraints did not cover the territory the model wandered into, or your format instruction did not match the shape you actually needed. Add one constraint or paste one example and run it again.

The first draft is 80% there. Not perfect. Prompts are not magic. But 80% on the first pass, with one iteration to tune, is the bar. If you are consistently at 30%, the context block is the problem.

When it breaks

Vague task with no format. "Help me think through our pricing strategy" with no deliverable or structure gets an essay on pricing frameworks you already know. Fix: specify what you want back. "List three pricing models worth considering, with one tradeoff for each. Use a table."

No business context. The model does not know you are bootstrapped, that your buyer is a contractor not a CTO, that $4M in revenue is the floor not the ceiling. Without context, it writes for the median company. Fix: add three bullets of context before the task. Industry, company size, audience, and what decision this feeds.

Skipped the constraints. The output is technically correct but stylistically wrong, full of hedging, corporate boilerplate, or a tone that sounds nothing like you. Fix: one specific constraint handles it. "No hedging language." "Write like a founder talking to another founder." "No corporate speak" is the sentence that earns its place in almost every prompt.

Starting over instead of appending. If the first output is close but not right, do not delete the prompt and rewrite from scratch. Add one line. "Make the recommended next step more direct." "Cut to five bullets." "You missed the action Sarah owns in paragraph three." One correction is always cheaper than a new brief, and it teaches you what your original prompt was missing.

Level up

Once you have the briefing structure running, the next move is making it persistent so you stop re-entering context in every session. Your first 30 minutes with Claude covers Projects, where you write your business context once and every conversation inherits it.

If you want to understand why any of this works at the model level, how AI works in plain English walks through the prediction mechanics without the jargon.

When you are ready to move past the chat window entirely, Claude Code tips for non-technical CEOs is the intermediate next step. That is where prompts become persistent, file-aware, and automated. The briefing habit you build here carries directly into that work.

For now: save the template. Pick one task you do this week. Run it with all five parts. The output will be different enough from what you have been getting that you will not need convincing.