Claude Code Adds a Prompt Audit for Outdated CLAUDE.md Patterns
Claude CodeClaude Code 2.1.283 added /doctor prompt-audit (also reachable as /checkup prompt-audit), a built-in check that scans a project's CLAUDE.md files, skills, agents, and commands for prompting patterns written for older Claude models. The same release improved the audit's output so stale file paths, stale commands, and instruction files that contradict each other lead the report, while thinking-related keywords that Claude Code's own documentation calls for are correctly left alone. The feature targets a specific problem: instructions written for older, less literal models, such as "verify twice" or heavy scratchpad scaffolding, can actively degrade behavior on current frontier models that already follow instructions closely.
Key Takeaways
/doctor prompt-audittreats prompt hygiene like a lint check, scanning CLAUDE.md, skills, agents, and commands for patterns written for older models rather than the current one.- Old prompting habits can actively hurt current models, since verification rituals and heavy emphasis language that helped earlier models often just add verbosity and extra tool calls on frontier models.
- The audit is also reachable as
/checkup prompt-audit, giving it two discoverable entry points inside Claude Code's existing setup-checkup command. - The improved report leads with the most actionable findings, surfacing stale paths, stale commands, and contradicting instruction files ahead of lower-priority issues.
- The fix also reduces false positives, since thinking-related keywords that Claude Code's own docs recommend are now correctly excluded from the stale-pattern flags.
- This formalizes a practice early adopters were already doing manually, as independent prompt engineers had been running similar audits against their own CLAUDE.md setups before it became a built-in command.
Sources & Mentions
3 external resources covering this update
Auditing Your Own Instructions, Not Just Your Code
Anthropic added /doctor prompt-audit, also available as /checkup prompt-audit, to check a project's CLAUDE.md files, skills, agents, and commands for prompting patterns that were written for older model generations. The underlying idea, already used in Anthropic's own claude-api skill, is that instructions tuned for earlier, less literal models can actively hurt performance on current ones. Patterns like repeated "double-check your work" verification rituals, "CRITICAL: YOU MUST ALWAYS" emphasis boosters, fixed-step scratchpad scaffolding, and few-shot examples tuned to an older model's failure modes tend to make a frontier model more verbose or more prone to unnecessary tool calls, rather than more reliable. Running the audit surfaces these leftover patterns across a project's configuration so they can be cleaned up.
A More Useful Report
Alongside the new command, Anthropic also improved how the audit presents its findings: stale file paths, stale commands, and instruction files that contradict one another now lead the report, ahead of less urgent findings, so the most actionable issues surface first. The audit also now correctly leaves alone thinking-related keywords that Claude Code's own documentation recommends, rather than flagging them as outdated alongside genuinely stale patterns.