A Claude Code skill that turns your messy AI-built project into a diagnosis report a non-developer can actually read. What is wrong → where to fix it → what to approve first. It only changes what you approve, and every change comes with a one-line undo.
Language: Korean conversations receive Korean reports; English conversations receive English reports. Machine-readable markers stay in Korean by design. Docs: English (this page) · 한국어 README
3-minute demo — see a real diagnosis report · Measurement records (EVALS)
English report example — the report explains the next safe action before anything changes.
Current version: v2.8.0 — Korean and English diagnosis reports, all modes + returning-patient features (grade trend "May D → today C", treated-area follow-up, one prescription per visit, checkup interval guidance), pre-execution change preview, write-boundary auto check, reports as Markdown / HTML ("clinical" report design) / PDF (auto-converted via your local browser) / Word, plus an HTML report security verifier (allowlist-based tag / external-resource / secret-leak machine checks). Changelog: CHANGELOG · License: MIT
| Mode | When | What you get |
|---|---|---|
Checkup checkup |
"Something feels wrong with this project" | Diagnosis against a fixed catalog (22 base checks) → treat only approved items → one-line undo guaranteed |
Deep checkup checkup --deep |
You want to look deeper | + git history hotspots ("this file changed 6 times"), dependency review, AI-artifact questionnaire |
Intro chart intro |
"I don't even know what this is" / handover | Project map (5 key files), how to run it, where not to touch |
Pivot pivot |
"Should I change direction?" | Compare paths → keep / modify / drop classification → milestones that still run if you stop |
Release check release-check |
Right before sharing / shipping | Secrets (including git history) and PII scan → pass / hold verdict |
Every report explains each finding with "What does it mean? / Where? / If fixed? / Approval command" in English reports, with the matching plain-language Korean labels in Korean reports.
Famous open-source projects, actually checked up — every finding re-verified against the code before publication:
- left-pad — the 11 lines that broke the internet · grade 🟢 B, ✅ ready to ship (English report)
- Moment.js — checkup of a retired giant · grade 🔴 D, size guard engaged honestly (English report)
- colors.js — the checkup that walked into a crime scene · grade 🟡 C, plus an out-of-catalog caution on the 2022 sabotage code (English report)
- Reactor project — a real D→C treatment story (Korean) · request — autopsy of an npm legend (Korean)
- Your project content is sent to Claude (Anthropic) servers — that is how Claude Code works. Check your company's AI policy before running this on confidential code.
- Cost: checkups consume Claude tokens.
--deepcosts several times more. - Scope and disclaimer: only known patterns defined in the diagnosis catalog are checked. Security vulnerability analysis and legal review are out of scope. Final judgment and responsibility stay with you.
- This is not a static analyzer: it is a diagnostic procedure (instructions) that Claude follows. Results can vary with model versions and project structure (details: limits of measurement).
Option A — one line (clone + install):
Windows (PowerShell):
git clone https://github.com/Ps-Neko/project-doctor.git; cd project-doctor; powershell -ExecutionPolicy Bypass -File install.ps1macOS / Linux:
git clone https://github.com/Ps-Neko/project-doctor.git && cd project-doctor && bash install.shThe script copies skills/project-doctor into ~/.claude/skills/, then verifies the install (SKILL.md presence + version). Safe to re-run — updating uses the same command. Restart Claude Code and /project-doctor is recognized.
- Check current install state without installing:
powershell -File install.ps1 -Check(macOS/Linux:bash install.sh --check)
Option B — Claude Code plugin (experimental):
/plugin marketplace add Ps-Neko/project-doctor
/plugin install project-doctor@project-doctor
Manual install (no script)
New-Item -ItemType Directory -Force ~/.claude/skills
Copy-Item -Recurse -Force skills/project-doctor ~/.claude/skills/macOS/Linux: mkdir -p ~/.claude/skills && cp -r skills/project-doctor ~/.claude/skills/
When updating manually, delete the old folder first so stale files do not survive: Windows
Remove-Item -Recurse -Force ~/.claude/skills/project-doctor, macOS/Linuxrm -rf ~/.claude/skills/project-doctor, then copy again. The install scripts do this for you.
/project-doctor # not sure where to start — intake will guide you
/project-doctor checkup "<path>" # health checkup (quote paths with spaces/Korean)
/project-doctor checkup --deep # deep checkup
/project-doctor intro # intro chart (explanation / handover)
/project-doctor pivot "<goal>" # pivot plan
/project-doctor release-check # pre-release check
The report comes first. You choose whether to approve a proposed change, one at a time.
- Diagnosis is read-only — nothing changes before the report is out
- Only approved items run — one at a time, with a "what will change" preview first
- Always undoable — checkpoint (git or backup) before execution, one-line undo after; if it cannot be undone, it does not run
- It remembers — approved treatments are recorded in
.project-doctor/, so the next checkup can tell you whether things actually got better
skills/project-doctor/ ← the skill that gets installed (SKILL.md + 8 reference docs)
tests/ ← quality tooling: scoring scripts + pytest
tests/fixtures/ ← test projects with planted problems + answer keys
SPEC.md / PLAN.md ← the skill's spec and build log
All secrets/PII inside
tests/fixtures/are fake by design — planted samples for measuring detection (tests/fixtures/README.md).
- Detection rate: internal fixture baseline — 3 consecutive independent-session runs at 100% against a 14-item answer key (pass bar: worst-of-3 ≥ 80%). Formal fresh 3-run measurement dates to v1.0.0; later versions keep catalog/scorer/fixtures unchanged and pass CI regression on every commit (limits).
- English reports (v2.8.0): measured on 2 fresh official English runs (+1 pilot) plus 1 Korean regression run, all passing format 0 / score 14-of-14 / false-positives 0 / location 5-of-5 — a smaller sample than the 3-run standard; model-generated reports are non-deterministic, so a small-sample pass can still be lucky (limits).
- Secret detection: internal baseline — fake-key samples, 3 consecutive 100% with 0 false positives (pass bar: 100% — one miss is an incident).
- All measurement is machine comparison of report IDs against answer keys (EXPECTED.md), not human impressions; reporting an ID not in the key counts as an automatic fail.
- Know this: scoring is ID-match based, the procedure is AI-followed instructions, and results can change when models change — which is why re-measurement is scheduled per quarter / per major model update (EVALS).