instructions-tuning

Use when creating, editing, splitting or auditing any file an agent reads to decide how to behave — CLAUDE.md, AGENTS.md, SKILL.md, agent definitions, .claude/rules, text a hook injects, system prompts — or when an instruction keeps getting ignored, misapplied, or contradicts another one. Also when one of our own hooks, scripts or agent tools misfires or stays silent when it should fire. Trigger words: 'инструкции', 'правило', 'правь CLAUDE.md', 'поправь скилл', 'мета-файл', 'хук не сработал'. Diagnoses why an instruction fails, picks the form that fixes it, and keeps the surface lean for current Claude models.