context-auditlisted

# Context Audit <context> Context files (CLAUDE.md, AGENTS.md, .claude/rules/) are always-on prompt context. Research baseline: frontier models follow ~150-200 instructions with reasonable consistency; Claude Code's system prompt consumes ~50 slots, leaving ~100-150 for user content. Adherence degrades linearly with instruction count (affecting ALL instructions uniformly), and files that impose unnecessary requirements can reduce task success versus no context file at all. The most effective files are under 100 lines with high directive density. Output is concrete fix proposals: exact changes with rationale, budget impact, and confidence. Deterministic checks (file existence, import traversal, symlinks) are labeled separately from heuristic judgment (directive counting, generic-advice detection). The audit prunes and grows in one pass: it removes what costs more than it earns, then proposes additions financed by what the pruning freed. </context> <constraints> - Never auto-commit; never delete content without presenting it to the user first - Label every finding deterministic vs heuristic; use `~N` for heuristic counts, not false precision - Never emit a letter grade or an aggregate quality score. A directive count is a measurement and is reported; a grade is a judgment dressed as a number and varies run to run. The health-check table and the budget band carry the summary - A file with no findings produces a report with no proposals, and that is a complete result. Never pr