chat-audit

原文涵盖的主题

• Chat audit
• Non-negotiables
• Route
• Tools
• Configuration
• Cost

Reading past sessions to improve how the work is done — not to summarize what happened. The failure this skill exists to prevent: a request like "re-read the last chats and find where things could be faster" produces two vague bullet points. Not because the sessions lack material — because nothing defined what "better" means, what will be done with a finding, or which sessions matter. An audit without a purpose returns platitudes. Establish the purpose first, always. 1. Never skip intake. Even when the request looks clear. modes/intake.md — one pass, then work. 2. Never read raw transcripts with your own eyes. They run to gigabytes. The .claude/skills/chat-audit/lib/ scripts turn them into a small structured slice; agents read the slice. Reading raw .jsonl by hand burns the budget and finds less. 3. Every finding carries an anchor — session id + timestamp + quote. An agent's report is a claim until the anchor is checked. Findings without anchors are dropped, not softened. 4. A finding that a rule already covers is a different, stronger finding. "There is no rule for X" and "the rule for X exists and did not fire" demand opposite fixes. modes/recon.md runs before analysis so you can

这里只展示节选，请以完整原文和当前要求为准。