The Anatomy of a Claude Prompt

SITUATION
[The facts, with numbers. What changed, over what period, and what did NOT change.]

QUESTION
[One open question. "What are the most likely causes?" Not "X is better, right?"]

ROLE
Answer as a [role, a few words].

SKILL
/[skill]  [One skill, only for a repeated or format-bound task. Delete this line for creative work.]

RULES (three, no more)
1. [rule]
2. [rule]
3. [rule]

BEFORE YOU ANSWER
Ask me about anything ambiguous first. Do not guess.

SOURCES
Give me a link I can open for every factual claim. Mark anything you could not verify.

Check the draft against each of these ONE AT A TIME, in this order.
After each one, tell me what you changed, or say "no change needed".
Then give me the revised version.

1. [constraint]
2. [constraint]
3. [constraint]
4. [constraint]

You don't need all seven parts every time. You need to know which one you skipped.

The same question, twice

Here is what I used to send:

My API is slow lately. How do I fix it?

That comes back with everything that can make an API slow. Add caching, check your indexes, look for N+1 queries, scale up the box. All correct, none of it about my problem.

Here is the same problem with the situation filled in:

Node 20 / Express API, ~40 req/s. p95 went from 180ms to 2.3s over 9 days,
gradually, not in a step. No deploys in that window. Mongo Atlas M10, main
collection grew 400k -> 3.1M docs in the same window. Memory flat ~380MB,
CPU under 35%.

What are the most likely causes, ordered by how cheap they are to rule out?
Ask me about anything ambiguous first.

Same model, same problem, and now I get a shortlist I can go test. "No deploys, memory flat, CPU flat" deleted most of the first answer before it got written.

The seven parts

1. Situation

Facts and numbers, plus one thing that did NOT change. "No deploys, memory flat, CPU flat" kills three wrong answers before they get written. Most people only describe what broke, and then they get a list of everything that could possibly break.

2. Question

Open, never yes/no. Ask "X is better, right?" and you get agreement. Ask "where does each option break first?" and you get a real answer.

3. Role

A job someone actually has, in about four words. "A backend engineer who has run this at scale" works. "A world-class 10x expert" does nothing. The job sets the vocabulary. The superlatives don't.

4. Skill

One skill when the task has a shape you repeat: a weekly report, a commit message, a review pass. None when you want ideas, because a skill narrows the output and that is the opposite of what you need. Never two at once. They fight each other and you get the overlap.

5. Rules, three max

Every rule has to be checkable by eye. "Max 150 words" is checkable. "Be concise but thorough" is not, and the model can't check it either. Past three, they start cancelling each other out.

6. Before you answer

One line: "Ask me about anything ambiguous first. Do not guess." Half my prompts get fixed right here, before I read 900 words built on the wrong assumption.

7. Sources

Ask for a link per claim, then add "mark anything you could not verify". The second half is the part that matters. Without it you get citations on the easy parts and silence on the hard ones.

Two examples, filled in

A decision:

SITUATION
Solo founder. One product live, 73 active users, no revenue yet. About 10
focused hours a week. Next three months I can only do one of these two.

QUESTION
Ship paid tiers on what already exists, or start the second product?
Where does each path break first?

ROLE
Answer as an operator who has run a one-person product business.

RULES
1. No outside funding, no hires.
2. Give me the week-4 signal that tells me it's working.
3. Pick one at the end. No "it depends".

BEFORE YOU ANSWER
Ask me about anything ambiguous first. Do not guess.

Nothing technical about this one:

SITUATION
A client is 3 weeks late on a 1,200 EUR invoice. Good relationship, first
time this has happened, I want to keep working with them. I sent one polite
reminder 8 days ago and got no reply.

QUESTION
What are my options for the next message, and what does each one cost me
in the relationship?

ROLE
Answer as a freelancer who has chased late invoices for years.

RULES
1. Max 120 words for the actual message.
2. No legal threats.
3. Give me two versions, one softer and one firmer.

The role changes, the rules change. The skeleton doesn't.

Before you hit send

• One number in the situation, and one thing that didn't change
• The question can't be answered with "yes"
• The role is a real job
• One skill or zero, never two
• Three rules, all checkable by eye
• You told it to ask before guessing

When to skip most of it

Brainstorming. Keep the situation and the question, drop the rest. Rules kill range.

Quick lookups. Two parts is the whole prompt. You don't need a role to ask for syntax.

In a codebase with an agent. The repo is your situation and the tests are your rules. Keep the open question and the "ask me first" line. That is what stops it from building the wrong thing for twenty minutes.

The second run

"Make it better" gets you a different draft, not a better one. Send your constraints one at a time and make it report after each one. The "no change needed" line is there so a good draft doesn't get wrecked by four rounds of edits nobody asked for.