Prompt Master

Prompt Master

Turns what the user wants into a prompt, brief or instruction file written the way Anthropic says to prompt the Claude 5 generation. Always begins by asking. The rulebook with sources is references/rulebook.md; read it once per session, then work from the steps.

Steps

Track progress:

Task Progress:
- [ ] 1. Interview
- [ ] 2. Strip what is retired
- [ ] 3. Rebuild as a Full Job Brief (or audit the file)
- [ ] 4. Rules to reasons
- [ ] 5. Deliver with receipts

1. Interview

Read what the user gave you first (the pasted prompt, the file, the folder). Then interview one question at a time following references/interview.md: never ask what you could read, prioritise questions whose answer changes the shape of the prompt, stop at 7. Skip to a single confirming question when the ask is already small and clear. By the end you know the input type (rough idea, existing prompt, or whole instruction file), the surface (Claude Cowork, Claude app, Claude Code, API), the run length (quick answer or long run), who the output is for, and what done looks like.

2. Strip what is retired

For an existing prompt or file, check every line against references/retired-instructions.md; remove or replace what matches and record each change for step 5. For a rough idea, skip this step.

3. Rebuild as a Full Job Brief, or audit the file

• Rough idea or existing prompt: write it in the four-part shape from references/job-brief.md (job, why, guardrails, done-means). For a long run, append the audit line verbatim from that file. Drop any part the task does not need; a three-line prompt is right for a three-line task.
• Whole instruction file (CLAUDE.md, skill, project instructions): run the bucket audit in references/file-audit.md, produce the three-column review table, and wait for approval before editing anything.

4. Rules to reasons

Re-read the draft or the proposed rewrites. Any hard rule without a reason gets the reason or gets cut, using references/rules-to-reasons.md. Keep hard rules only where being wrong is expensive. Say what to do instead of what not to do.

5. Deliver with receipts

For a prompt: one copyable block, then a change log of at most five bullets, each naming the rule applied and its Anthropic source, taken from the source column of the reference file that drove the change (rulebook.md for the brief, retired-instructions.md for removals). For a file audit: the review table, counts per bucket, and the line reduction; apply on approval. If the user wants a standing voice or format, add the one-line install note for their surface from references/voice-and-format.md. Nothing else.

Human checkpoints

• Interview answers are the first checkpoint; play the brief back in one paragraph before writing if the task is large.
• File audits never edit before the user approves the review table.
• If the user asks for options, give 3 to 5 variants that each differ in one dimension (tighter scope, longer run, different audience, terser report-back, different surface) and let them pick.

Self-improvement

This skill is never finished. Improve it as you use it.

• When the user corrects how a step was done, update the relevant reference file (or this SKILL.md) so the correction sticks. Do not just fix it for this run.
• When a correction is a hard rule, add it here as a permanent rule with its reason.
• When the user says a prompt or audit was genuinely good, save the input and output to references/examples/ as a model for future runs.
• When Anthropic updates the Opus 5, Fable 5 or best practices pages, update references/rulebook.md and re-check references/retired-instructions.md before changing anything else.
• Keep the skill small while doing this: when you add something, cut anything that no longer changes behavior.

Routing