Use when auditing the instruction context an agent actually receives — on explicit request ("audit my context/CLAUDE.md/system prompt", "context audit", "prune my instructions"), when a cross-layer instruction conflict is detected mid-task (two active layers direct incompatible behavior), or after a model-generation upgrade invalidates guardrails written for a weaker model. Do NOT fire for auditing the prose quality of one document (ordinary editing), for designing NEW tool/agent interfaces (that is interface craft doctrine, not context hygiene), for pre-work recon on a task brief (recon owns the territory; this skill audits the map), or for tuning a prompt to improve the…
