Linters check the file. Analytics grade your sessions. The doctor cross-examines one against the other — and cites receipts.

Quickstart

As a Claude Code plugin (recommended):

/plugin marketplace add agent-clinic/claude-md-doctor

then install claude-md-doctor from the /plugin menu. Or via the skills.sh CLI:

npx skills add agent-clinic/claude-md-doctor

Or bare: copy skills/claude-md-doctor/ into ~/.claude/skills/.

Then, in any repo, just ask — "give my CLAUDE.md a checkup" — or invoke directly: /claude-md-doctor:claude-md-doctor (bare install: /claude-md-doctor). The report lands in .claude-md-doctor/report.html plus machine-readable report.json.

Requires Python 3.9+ (standard library only). Everything runs locally; nothing leaves your machine.

What the exam covers

  • Vitals — effective size vs the official guidance ("target under 200 lines per CLAUDE.md file" — Claude Code memory docs), estimated token cost per session, structure, and pathology markers: stock /init boilerplate never pruned, emphasis saturation, changelog accretion.
  • Records check — does everything the file points at exist? Dead file paths, paths from a teammate's machine, @imports that don't resolve, pnpm/make commands with no matching script, .claude/rules/ scopes that match zero files.
  • Checkable claims — countable assertions ("2,100 tests across 180 files", "9 UI components; no dialog") verified against the repo. Inlined numbers rot; the doctor catches them.
  • The report — a single self-contained HTML page: chart grade, chief complaint, per-finding evidence, the session-adherence History table, and concrete prescriptions, each footnoted with the official doc or study behind it (the evidence base lives in docs/RESEARCH.md).

It understands the real memory surface: CLAUDE.md, .claude/CLAUDE.md, CLAUDE.local.md, nested files, .claude/rules/*.md (with paths: scopes), @imports (depth 4, backtick-aware), claudeMdExcludes, ancestor directories — and it treats the pointer-to-AGENTS.md pattern as healthy, examining the target, while flagging the broken variant (pointer text without @, which Claude Code never actually loads). A repo with an AGENTS.md but no CLAUDE.md at all gets the doctor's simplest prescription: the official one-line pointer, so Claude Code stops loading nothing.

The backtest — check if CLAUDE.md actually works in your sessions

Your own Claude Code session transcripts (~/.claude/projects/…) already record whether past sessions actually followed each rule in your CLAUDE.md. The doctor decomposes the file into rules and replays them against that history — per rule, a verdict with receipts:

Rule Opportunities Compliance Verdict
Never import the legacy API types 12 100% healthy
Run verify before you finish 2 0% ignored
Never hardcode a colour 0 — inert

Behind every number: matched excerpts, and for finish-ordering rules a session-timeline strip showing exactly what ran after the last edit. Two ideas drive the verdicts (full taxonomy). Every rule gets an enforcement class — the cheapest reliable detector:

Class Detector Binds
hook gate over tool calls (commands, edits, orderings) — prevents the agent
linter/test static analysis over the code itself every agent and every human
judge LLM audit, post-hoc, with a stated reliability ceiling audit only

~70% of real-world directives land in the first two — laws waiting to be passed. And every violation is triaged by cause, because the cause picks the medicine:

Cause What happened Medicine
defiance-proven the agent echoed the rule, then broke it block-mode gate — the reminder already lost
defiance violated in fresh context warn-hook, then block
dilution drowned late in a heavy session slim the file, move the rule to point-of-use
absence non-root rule lost to compaction re-inject; never block

The arming ladder (reminder → warn → block) is set per rule from its own violation forensics, and review-then-arm hook proposals are written to the exam folder — nothing is ever installed automatically. Every checkup also emits a share-safe card (grade, hearts, doctor's note — aggregates only, never a string from your repo) and a claude-md-health.svg badge for your README. Matcher fires are sample-verified before they count, because matchers have bugs; unverified results are banner-labeled provisional. Research shows agents silently skip mandated steps while outputs still pass checks; only behavioral evidence catches that — and it's free, sitting in your transcript history.

Run it on your own repo: the rules you'd bet on being followed are rarely the ones that are.

No CLAUDE.md? The doctor writes your chart

Most repos have no memory file at all (18 of the 20 on our own machine). But their session transcripts already contain the unwritten rulebook, and the same engine that backtests rules can run in reverse — mine the history, then validate the checkable candidates against it:

Signal Example Becomes
repeated corrections "no, use pnpm not npm" typed in 3 sessions a rule
failed → fixed pairs npm test fails, pnpm test works, again a rule (often a hook)
re-discovery agent reads package.json at every session start a fact, stated once
permission denials you rejected git push twice a "never" rule
repeated preambles the same context paragraph pasted each session a fact

Grouped signals survive only with recurrence (≥2 sessions or ≥3 occurrences; re-discovery needs 3 distinct sessions) and carry recency flags — a preference the repo moved past is marked stale for the judge pass to decline. Corrections reach the judge ungated (wording varies too much to group), deduped and capped, and are judged hardest. Each accepted mechanically-checkable rule is then replayed through the backtest for precise counts. The result is PROPOSED-CLAUDE.md: a lean draft where every line carries its receipt as an HTML comment (stripped at load, so it costs the adopter nothing), hook-class rules arrive as review-then-arm guard proposals ("born mechanized"), and the report shows the re-discovery tax your sessions have been paying. The draft is held to the same 200-line vitals this tool grades everyone else on — the generator refuses to prescribe the disease it diagnoses. Nothing is installed and no CLAUDE.md is written for you — the draft lands in the exam folder (.claude-md-doctor/), and adoption is your move. Repos that do have a CLAUDE.md get the same mining as a gap analysis: rules you keep dictating by hand that the file never says.

FAQ

Why does Claude ignore my CLAUDE.md? Usually one of three causes, and they need different medicine: defiance (the rule was in context — sometimes literally echoed — and broken anyway), dilution (the rule drowned late in a heavy session), or absence (a non-root rule lost to compaction). The research says this is normal, not user error: agents violate 57.5% of preference rules even when memory retrieves them, and perform about half the steps their own instruction files mandate (docs/RESEARCH.md). The backtest tells you which cause is yours, with the transcript as receipts.

Can it write my CLAUDE.md for me? Yes — from evidence, not from templates. If your repo has no memory file, the exam switches to intake mode: it mines your local session transcripts for recurring corrections, failed→fixed commands, re-discovered facts, and denials, replays the checkable candidates against that same history, and drafts PROPOSED-CLAUDE.md with a receipt on every line. Unlike /init, which reads your file tree, this reads your behavior — it only proposes rules you have demonstrably needed, with recurrence gates and staleness flags to keep one-off taste out.

How do I audit my CLAUDE.md or AGENTS.md? Install the skill (Quickstart above), then ask in any session: "give my CLAUDE.md a checkup" — audit, review, improve, and lint requests all route to the same exam. The report lands in .claude-md-doctor/report.html.

How long should a CLAUDE.md be? The official guidance says "target under 200 lines per CLAUDE.md file" (memory docs). The doctor measures your effective loaded size — imports resolved, fences and comments handled — against that number, and estimates what the file costs in tokens per session. Length itself is the weakest signal, though; the strong causal evidence is about pruning content that doesn't change behavior.

How is this different from /doctor or the official claude-md-management plugin? Those score or trim the file, statically. This joins the file to your behavior: it replays every rule against your real session transcripts and shows per-rule compliance with cause-of-failure triage. Linters check the file; analytics grade your sessions; the backtest is the join.

Does my session data leave my machine? No. Everything runs locally, stdlib Python only, no telemetry. The share card and badge are aggregates-only by construction — never a string from your repo or transcripts — and there's a test asserting exactly that.

Honesty policy

Every prescription carries an evidence tier — official doc, controlled study, corpus study, or plainly-labeled heuristic — and the report states the tensions in the research instead of hiding them (e.g., the one factorial study found no structural effect of file size in its tested range, while content pruning has strong causal backing). See docs/RESEARCH.md for the full evidence base — 40+ primary-verified sources.

Status

The full exam works end to end: static checks + session backtest + cause triage + enforcement compilation + generative intake mode (no CLAUDE.md → mine sessions → drafted chart with receipts) + share card/badge + verified report. Tested (python3 -m unittest discover -s tests), calibrated against real-world gold-standard files (python3 fixtures/fetch.py), and dogfooded on a real repo — including a clean-context validation run, where a fresh agent guided only by the skill's own instructions completed every stage, caught a matcher bug via the built-in verification loop, and found two real problems the authors had missed. Part of agent-clinic — checkups for your agent's config files. Issues and PRs welcome.

License

MIT. claude-md-doctor is an independent open-source project, not affiliated with Anthropic. CLAUDE.md and Claude Code are products of Anthropic, PBC.