Skill lens

<!-- LEVEL 1 — Read this first (15 lines) -->

What	Action
Goal	Find installed skills that NEVER auto-invoke because they're missing from the hot-trigger list
Input	~/.claude/skills/ (real) + ~/.claude/CLAUDE.md hot-trigger list
Output	Coverage report: ✅ Covered · ⚠️ Blind Spots · 👻 Ghosts + coverage %
Optional	--apply flag auto-adds top 10 blind spots to hot-trigger list
Manual trigger	"skill-lens" / "check my skills" / "检查盲区" / "有哪些 skill 漏了"
Auto-trigger	After brain-sync completes

---

<!-- LEVEL 2 — Full pipeline starts here -->

Skill Lens v1.1 — "Your hot-trigger list has blind spots."

The Problem

You have 100+ skills installed. Only a fraction are in the hot-trigger list that actually gets auto-invoked. The rest are installed but invisible — they exist on disk but will never be called because:

1. The AI doesn't see them in the hot-trigger checklist
2. They're buried in the system reminder (144 skills = 200+ tokens)
3. You forgot you installed them

What It Does

Cross-references installed skills (~/.claude/skills/) against the hot-trigger list in ~/.claude/CLAUDE.md:

~/.claude/skills/              CLAUDE.md hot-trigger list
      │                                  │
      ▼                                  ▼
  installed skills                auto-invoked skills
      │                                  │
      └────────── cross-reference ───────┘
                     │
        ┌────────────┼────────────┐
        ▼            ▼            ▼
   ✅ Covered    ⚠️ Blind     👻 Ghosts
   in both      in skills    in hot list
                but NOT in   but NOT in
                hot list     installed

Quick Run

# Show report only
node ~/.claude/skills/skill-lens/lens.cjs

# Auto-apply top 10 to hot-trigger list (with backup)
node ~/.claude/skills/skill-lens/lens.cjs --apply

# Auto-apply top N
node ~/.claude/skills/skill-lens/lens.cjs --apply --top 5

Sample Output

## Skill Lens Report — 2026-06-07

📊 **Coverage: 25.7% (37/144)** ↑ 0.69% vs last run

✅ Covered:   37 skills
⚠️  Blind:    107 skills (installed but NOT auto-invoked)
👻 Ghosts:    0 skills (in hot list but NOT installed)

### ⚠️ Top 10 Blind Spots (by priority)
 1. baoyu-comic              (priority 75)
 2. baoyu-danger-gemini-web  (priority 75)
 3. baoyu-infographic        (priority 75)
 4. canary                   (priority 65)
 5. careful                  (priority 65)
 6. connect-chrome           (priority 65)
 ...

What the Script Does (for LLM understanding)

The lens.cjs script handles all the dirty work so you don't have to:

Step	What it does	Why
1. Extract	ls ~/.claude/skills/ filtered by SKILL.md existence	Real source of truth, not docs.md
2. Extract	Parse CLAUDE.md hot-trigger section, split multi-skill rows (**a / b**)	Avoid double-counting
3. Compare	Three-set: covered / blind / ghosts	Standard Venn diagram
4. Score	Priority = 50 (base) + 40 (own product) + 25 (baoyu-) + 15 (gstack)	Rank by importance
5. Apply (opt)	Backup CLAUDE.md → insert top N rows → update "If none of the {N} match"	Reversible
6. History	Append to ~/.claude/.cache/skill-lens-history.json	Track coverage over time

Pipeline (if you want to do it manually)

Step 1: Extract

# Source A: All installed skills (real)
ls ~/.claude/skills/ | grep -v "^_\|^\."

# Source B: Skills in hot-trigger list
sed -n '/Hot Trigger List/,/If none of the {N} match/p' ~/.claude/CLAUDE.md | \
  grep -oP '\*\*[a-z][a-z0-9_-]*[a-z0-9]\*\*' | sort -u

Step 2: Compare

Build three sets using comm:

• ✅ Covered: in BOTH
• ⚠️ Blind: in skills, NOT in hot list
• 👻 Ghosts: in hot list, NOT in skills

Step 3: Prioritize

Sort blind spots by:

1. User's own products (tool-eval, preflight, brain-sync, skill-lens, skill-creator)
2. baoyu-* high-frequency skills
3. gstack curated suite (browse, canary, careful, codex, etc.)
4. Everything else

Step 4: Apply (manual)

To add a blind spot to the hot-trigger list, edit ~/.claude/CLAUDE.md:

1. Find ## Skill Auto-Discovery — Hot Trigger List
2. Insert new rows before **If none of the {N} match:**
3. Update the {N} in the "If none of" line

Or just run: node lens.cjs --apply

History Tracking

Each run appends to ~/.claude/.cache/skill-lens-history.json:

[
  {"date": "2026-06-07", "total": 144, "covered": 36, "coverage": 25.0, "delta": null},
  {"date": "2026-06-07", "total": 144, "covered": 37, "coverage": 25.7, "delta": 0.69}
]

Use this to spot regressions (coverage drop = someone removed hot-trigger rows).

Rules

• Run after every brain-sync. Auto-triggered by brain-sync-complete event.
• Don't just list — prioritize. Every blind spot comes with a priority score.
• Backup before apply. The --apply mode auto-backs up to ~/.claude/.cache/skill-lens-backups/.
• Skip docs. Skills without SKILL.md are not counted (avoids framework/ repo false positives).
• Source of truth = ~/.claude/skills/. skill-directory.md is documentation, not config.

Changelog

v1.1.0 (2026-06-07)

• NEW: lens.cjs script with --apply auto-mode
• NEW: Coverage % metric with delta tracking
• NEW: History log at ~/.claude/.cache/skill-lens-history.json
• NEW: Priority scoring (own products +40, baoyu- +25, gstack +15)
• NEW: Multi-skill row handling (**a / b** → 2 skills)
• NEW: Skip skills without SKILL.md (framework/ false positive filter)
• FIX: Removed hardcoded 137/15 numbers (now dynamic)
• FIX: Tags expanded: discovery, audit, gap-analysis
• FIX: argument-hint moved to dedicated frontmatter field

v1.0.0 (2026-05-XX)

• Initial release: markdown instructions, manual bash extraction