Claude Code 2.1.283 adds /doctor prompt-audit: it reads your CLAUDE.md, skills, agents and commands and reports stale paths, stale commands, contradicting instruction files, and wording written for older models. You are on 2.1.282. Today: run it, read it, keep the top findings.
The 9-23 audit found stale CLAUDE.md and README lines by hand (batch 6) and memory that still says 'never PDF'. Every session loads your 200-line CLAUDE.md plus sixty-odd skill descriptions, so a stale instruction is paid for on every turn and occasionally obeyed. This is the same check, run by the tool, for free.
Instruction files age in three ways. Paths move (the ~/Documents to ~/Claude move left dozens of dead references). Rules get superseded but not deleted (paper-first on 8-31 versus skills that still write HTML cards). And prompting habits written for older models persist: emphatic capitals, 'think step by step', long lists of don'ts. Newer models follow plainer prose better and the old scaffolding costs tokens.
prompt-audit is non-interactive. It inventories every instruction file it can see from the current directory, checks provenance, scans for dated patterns, and produces a report with file:line, the pattern, why it is obsolete for the current model, and a confidence. It proposes a diff; it edits nothing unless you ask.
Read it like the 9-23 findings table: high-confidence stale paths and contradictions first, style suggestions last. Some 'old-model' patterns in your CLAUDE.md are deliberate (the load-bearing markers, the IMPORTANT lines) and the audit keeps documented Claude Code keywords. Treat its diff as a proposal to the same standard as any other.
Update, then run the audit from your workspace root so it sees both CLAUDE.md files.
claude update && claude --version
cd /Users/tom/Claude && claude
# inside the session:
/doctor prompt-auditInstruction files rot like code; now there is a linter for them.