instructions-tuning

原文涵盖的主题

• Instructions Tuning
• Where to look
• Process
• 1. Diagnose
• 2. Pick the form
• 3. Write it
• 4. Budget-check
• 5. Apply and record

Editing the files an agent reads to decide how to behave. Two facts frame every edit. Current Claude models follow instructions closely and literally, and have good judgement. Most failures now come from too much instruction, not too little: a rule applied exactly where it is wrong, rituals that make the model write more and repeat tool calls, two layers saying opposite things. Anthropic removed over 80% of Claude Code's own system prompt for the Claude 5 generation with no loss on their evals. So the first fix to consider is a deletion, or rewriting a rule into the principle behind it; adding text comes second. A rule that keeps failing usually has the wrong form, not the wrong words. Match the form to the way it fails. | Situation | Read | |---|---| | Fixing or writing one instruction — a CLAUDE.md line, a rule, a skill section | this file | | Writing a skill, restructuring or splitting one, fixing its description | skill-files.md | | Auditing a whole surface — a project's CLAUDE.md, rules, skills — for dated patterns | audit.md | | One of our own hooks, scripts or tools misfires or stays silent | tool-holes.md | | Writing or fixing a bash hook | bash-hooks.md | | Writing or fixi

这里只展示节选，请以完整原文和当前要求为准。