YAGNI Prompt Audit Kit: Rubric and 30 Before-After Rewrites

A prompt that opens with a greeting, a persona and a plea to be thorough pays for those words on every call and gets a vaguer answer. This kit gives you a rubric, a blocklist and a script to find the dead weight, plus 30 worked rewrites that cut 3,058 estimated tokens to 1,726 while adding an output shape and a check.

What You Get

• Eight-criteria rubric. Single job, needed context, testable constraints, output shape, no filler, no generic advice, stop condition and verification, each scored 0 to 2 with fixes and four action bands. Available as markdown and JSON.
• Filler-phrase blocklist. 122 phrases in 10 groups, from greetings and persona claims to effort pleas and generic advice. Plain text, so you can add your own.
• Audit script. Reads a prompt or file, estimates tokens, lists blocklist hits by line, flags contradictions and unbounded tasks, and scores the six criteria it can judge from the text.
• 30 before and after rewrites. Ten coding prompts, five writing, four data, five agent system prompts and six CLAUDE.md sections, each with token counts, scores and a note on what was cut and added.
• rewrites.json. The same 30 pairs as data with computed token counts, reductions and scores.
• CLAUDE.md diet checklist. A six-step process with a keep test, a table of cut-or-move decisions and signs that a file has bloated. Usually removes 40% to 60% of a file.
• Review template. A copy-and-paste form that records the measure, the score, the rewrite, the evidence and the decision, with a cost example.
• Nine automated tests. Including a check that every shipped rewrite scores higher than its original and has no filler left.

How to Use

1. Run python3 tools/audit_prompt.py on a prompt or your CLAUDE.md.
2. Score R1 and R2 by hand and read the band for the total.
3. Rewrite with the five-step method in RUBRIC.md, deleting the lines the tool lists first.
4. Compare against the closest of the 30 rewrites to see the pattern.
5. Test old and new prompts on three real inputs and keep the one that holds up.
6. Fill in the review template and commit the token counts with the change.

Who This Is For

• Developers whose CLAUDE.md or AGENTS.md has grown past what anyone reads.
• Teams running agents in production who pay for system-prompt tokens on every call.
• Prompt authors who want a repeatable review instead of a gut feel.
• Consultants auditing a client's prompt library.

Everything runs offline on Python's standard library, so you can audit private prompts safely. Browse more in the system prompt libraries collection, or try Prompt Bloat Analyzer.