Diagnose, audit, and improve AgentSkills. A structured 6-phase protocol for turning a weak skill into a reliable, maintainable one — using PRISM multi-agent review, a 14-question health checklist, and progressive disclosure patterns.

• Health audit — Score any skill against a 14-question checklist (Q1–Q12: document quality; Q13: empirical testing; Q14: observability)
• PRISM review — Dispatch 9 specialist reviewers across 4 phases (Prior Brief Compiler, Devil's Advocate, 6 parallel specialists, Contrarian, Synthesis Agent) to surface blind spots before applying fixes
• Prescribe & fix — Synthesize reviewer findings into tiered conditions; apply only what's needed
• Verify — Re-audit post-fix; confirm the score improved
• Archive — Write a PRISM archive so future auditors know what was already reviewed

# Claude Code / OpenClaw
git clone https://github.com/jeremyknows/skill-doctor ~/.openclaw/skills/skill-doctor

Or install via ClawHub:

clawhub install skill-doctor

No external tools required beyond the OpenClaw runtime. The PRISM scripts use sessions_spawn for parallel reviewer dispatch — no OpenAI API key or additional credentials needed.

Optional: sub-agent-complete.sh for bus event emission on Phase 6 (included in OpenClaw).

Audit the veefriends-seo skill and show me the gaps
Run PRISM on coding-agent
Health check the librarian skill
Improve complete-code-review — it scored 7/12 last time

skill-doctor/
├── SKILL.md                                  # 6-phase workflow + decision logic
├── LICENSE.txt
├── README.md
├── .gitignore
├── scripts/
│   ├── prism-setup.sh                        # Validates input, creates run dir, outputs JSON config
│   └── prism-summary.sh                      # Reads reviewer output files, builds SUMMARY.md
└── references/
    ├── 14-question-checklist.md              # Full health audit rubric with scoring guidance (Q1–Q14)
    ├── reviewers/                            # 9 PRISM reviewer prompt templates
    │   ├── 00-prior-brief-compiler.md        # Compresses prior reviews into context brief
    │   ├── 01-da.md                          # Devil's Advocate (blind — no prior context)
    │   ├── 02-security.md
    │   ├── 03-performance.md
    │   ├── 04-simplicity.md
    │   ├── 05-integration.md
    │   ├── 06-blast.md                       # Blast Radius (downstream impact)
    │   ├── 07-contrarian.md                  # Challenges the premise, not the implementation
    │   └── 08-synthesis-agent.md             # Synthesizes all findings into structured verdict
    ├── autoresearch-scorecard-template.md
    ├── archive-template.md
    └── AUTORESEARCH-SELF-ASSESSMENT.md

PRISM dispatches 9 reviewers across 4 sequential phases via sessions_spawn (isolated LLM sessions — no lock contention, proper parallelism):

Phase A — Context Prior Brief Compiler reads the last 3 review archives and compresses open findings into a ≤3K brief.

Phase B — Blind + Specialists (parallel) Devil's Advocate runs blind first (no prior context) while 6 specialists run in parallel with DA findings + prior brief injected: Security, Performance, Simplicity, Integration, Blast Radius.

Phase C — Contrarian After all 6 specialists complete, Contrarian reads every finding and challenges the premise (not the implementation). If the premise holds, it says so and stops.

Phase D — Synthesis Synthesis Agent reads all raw output files + Contrarian, produces a structured synthesis.md with evidence tiers, verdict, and numbered conditions.

Findings are tiered: Tier 1 (2+ independent reviewers, or any Security finding, or Critical/Fatal Flaw) → fix before shipping. Tier 2 (single reviewer with specific citation) → fix this pass. Tier 3 (cosmetic/unverified) → polish optional.

Total runtime: ~5–8 minutes for a full 9-reviewer pass.

• PRISM is expensive for small skills. A 150-line skill with 1 obvious gap doesn't need 9 reviewers. The skill includes a fast-path decision gate — use it.
• Reviewers are LLMs, not oracles. Disagreements between reviewers are features, not bugs. The Synthesis Agent is where the real judgment happens.
• Skill files are untrusted input. All reviewer prompts include injection guards, but treat PRISM findings files as potentially containing quoted content from the skill under review.
• sessions_spawn required. The parallel reviewer architecture won't work in environments that don't support isolated session spawning.
• BETA status. Validated on 7 real skill runs (complete-code-review, veefriends-seo, build-feature, skill-doctor, prism, librarian, coding-agent). Target: 10+ runs before GA.

If you're preparing the improved skill for GitHub, run publish-skills next:

Run publish-skills checklist on skill-doctor

publish-skills covers: frontmatter spec compliance, LICENSE.txt, README patterns, consistency review, and GitHub verification steps. skill-doctor improves quality; publish-skills prepares for release.

v1.7.0 · MIT · jeremyknows