dmminga1983/minga-prompt-audit

An audit skill for Claude prompts and skills: outdated techniques, conflicts, dead references, bloat, skill structure, security, clarity and triggering — reported line by line with a ready-to-apply diff.
Part of the Minga ecosystem. The skill checks CLAUDE.md, AGENTS.md, SKILL.md, rules, commands, subagents and system prompts for patterns that hold modern Claude models back: ALL-CAPS pressure, forced reasoning scaffolds, rules that contradict each other, references to files that don't exist, duplicates, broken skill frontmatter, leaked secrets and personal data, vague requirements, skills that never trigger. Nothing in your files changes until you say "apply".
Русская версия →
## How it differs from the built-in prompt-audit
| | /claude-api prompt-audit, /doctor prompt-audit | minga-prompt-audit |
| Where it runs | Claude Code CLI | claude.ai, Claude Code, any agent that supports skills |
| Check catalogue | not published | open, with IDs and examples (checks/) |
| Russian text | — | Russian pressure and persona patterns; abbreviations such as ГОСТ, СП, ЖК are not treated as caps |
| Deterministic scanner | — | scripts/scan.py, no dependencies, with tests |
| Protected lines | keep-list | and … blocks |
## What it checks (v1.1)
| Block | Checks |
| A. Outdated techniques | A1 caps pressure · A2 threats and rewards · A3 "think step by step" · A4 empty persona · A5 repeating a rule "to be safe" · A6 absolutes without a reason · A7 prohibitions only |
| B. Conflicts | B1 within a file · B2 across files · B3 format/language · B4 unclear priority |
| C. Dead references | C1 files and @imports · C2 commands and scripts · C3 tools and skills · C4 outdated model names |
| D. Bloat | D1 duplicates · D2 long examples · D3 the obvious · D4 oversized files |
| E. SKILL.md structure | E1 missing or broken frontmatter · E2 name not kebab-case or not matching the folder · E3 vague description · E4 description over ~1024 characters · E5 no step-by-step order · E6 links to missing skill files |
| F. Security | F1 secrets (AWS, GitHub, Slack, sk- keys, private keys, password = "...") · F2 personal data (phones, emails, ID numbers) · F3 obeying instructions from external content (prompt injection) · F4 dangerous actions without confirmation |
| G. Clarity | G1 vague requirement with no success criterion · G2 no output format · G3 no stop condition or point to ask the user · G4 unexplained jargon and abbreviations |
| H. Skill triggering | H1 no real trigger phrases (RU and EN) · H2 overlap with another skill in the same folder · H3 overly broad trigger · H4 no "when not to use" next to a similar skill |
Detailed criteria are in checks/ (in Russian), before/after examples in examples/.
Secrets and personal data are masked in every report: only the first 4 characters are shown.
## Install
claude.ai — on the repository page click Code → Download ZIP and upload the archive in Settings → Capabilities → Skills.