Skill Monitor

Closed-loop skill effectiveness monitoring. Reads session metrics, computes per-skill signals, identifies what's working and what needs improvement.

Inspired by the deploy-monitor-evaluate-improve feedback loop: skills get better over time instead of staying static.

Requirements

Requires .claude/session-metrics/metrics.jsonl from /session-scan. If no data: suggest running /session-scan first.

Usage

/skill-monitor                     # Dashboard: all skills
/skill-monitor --skill review      # Deep-dive on one skill
/skill-monitor --improve           # Generate improvement recommendations
/skill-monitor --window 30d        # Change comparison window (default: 7d)

What Main Context Does

Step 1: Parse Arguments

Extract from $ARGUMENTS:

• --skill NAME: Focus on one skill (e.g., review, plan, investigate)
• --improve: Spawn analysis agent for improvement recommendations
• --window PERIOD: Comparison window (7d, 30d, all; default: 7d)

Step 2: Load Metrics

Read .claude/session-metrics/metrics.jsonl. For each entry, extract the skill_effectiveness field (added by compute-metrics.py v2).

Filter by window period. Count sessions with and without skill usage.

If no skill_effectiveness data exists in metrics: "Metrics were computed before skill tracking was added. Run /session-scan --rescan to recompute."

OTel invocation_trigger (CC v2.1.126+): when compute-metrics.py ingests claude_code.skill_activated events, each invocation carries an invocation_trigger of "user-slash", "claude-proactive", or "nested-skill". If absent (older sessions), default to "unknown" — do NOT assume "user-slash".

Step 3: Compute Per-Skill Aggregates

For each skill found across all sessions, aggregate:

| Metric                  | Computation                                    |
|-------------------------|------------------------------------------------|
| Total invocations       | Sum of invocation_count across sessions        |
| Sessions used in        | Count of sessions containing this skill        |
| Action rate             | Weighted avg of per-session action_rate         |
| Avg post-errors         | Weighted avg of avg_post_errors                |
| Avg post-corrections    | Weighted avg of avg_post_corrections           |
| Outcome distribution    | Count of effective/friction/no_action/mixed    |
| Effectiveness score     | action_rate - (0.3 * avg_post_corrections)     |
| Adjusted score          | For analysis/check skills, use lower thresholds |
| Trigger distribution    | Counts of user-slash / claude-proactive / nested-skill / unknown |
| Proactive trigger rate  | claude-proactive / (user-slash + claude-proactive + nested-skill) |
| Auto-load gap           | Skills with 0 claude-proactive invocations across window |

Auto-load gap detection (CC v2.1.126+): Skills with auto-loaded behavior in their description (i.e., not disable-model-invocation: true) are EXPECTED to fire as claude-proactive. A skill that is ONLY ever invoked via user-slash is failing its description's routing intent. Flag any auto-loadable skill where proactive_trigger_rate == 0 over the window. This is the structural answer to the "zero skill auto-loading" gap from the 137-session analysis (see MEMORY.md). Confidence floor: only flag if total invocations >= 5 in window.

Skill type weighting: Analysis and check skills (verify, triage, perf, boundaries, pr-review, audit) have low action rates BY DESIGN — their success is "found issues" or "confirmed things pass". Apply adjusted thresholds:

Skill Type	Flag Threshold	Expected Action Rate
Execution (work, quick, full)	< 0.5	> 0.7
Analysis (perf, boundaries, audit, pr-review)	< 0.3	0.3-0.5
Check (verify, triage)	< 0.1	0.0-0.3
Knowledge (compound, learn, brief)	< 0.5	> 0.5

Also compute baseline friction (avg friction of sessions WITHOUT any skill usage) vs skill friction (avg friction of sessions WITH skill usage). Delta = skill_friction - baseline_friction. Negative delta = skills reduce friction (good).

Step 4: Display Dashboard

Dashboard mode (no --skill):

## Skill Effectiveness Dashboard (last {window})

Baseline friction (no skills): 0.32 | With skills: 0.18 | Delta: -0.14

| Skill           | Uses | Sessions | Slash/Proactive/Nested | Action% | Errors | Corr | Outcome   | Score |
|-----------------|------|----------|------------------------|---------|--------|------|-----------|-------|
| /phx:review     | 12   | 8        |    8 /  3 /  1         | 92%     | 0.5    | 0.1  | effective | 0.89  |
| /phx:plan       | 9    | 7        |    9 /  0 /  0         | 100%    | 0.2    | 0.0  | effective | 1.00  |
| /phx:investigate| 5    | 5        |    5 /  0 /  0         | 80%     | 1.2    | 0.4  | mixed     | 0.68  |

Skills needing attention:
- /phx:investigate (high post-errors)
- /phx:plan (auto-load gap — 0/9 proactive; description not routing)

Flag skills using type-adjusted thresholds (see weighting table above). Also flag if avg_post_corrections > 1 or outcome is predominantly "friction". Also flag auto-load gap: auto-loadable skills (without disable-model-invocation: true) with proactive_trigger_rate == 0 and total invocations >= 5. This is a description/routing problem — the skill exists but Claude isn't loading it on its own.

When displaying flagged skills, note if the flag is "expected" for the skill type (e.g., verify at 0.24 is normal for a check skill).

Skill deep-dive (--skill NAME):

Show per-session breakdown for that skill, including session IDs, dates, individual outcome signals, AND invocation_trigger per invocation. If a skill is dominated by user-slash triggers, surface which 1-3 description keywords might unlock proactive routing — cross-reference against the skill's current description in plugins/elixir-phoenix/skills/{name}/SKILL.md. If session reports exist in .claude/session-analysis/, reference them.

Step 5: Improvement Mode (--improve)

Spawn skill-effectiveness-analyzer agent:

Agent(subagent_type="skill-effectiveness-analyzer", model="sonnet", prompt="""
Analyze skill effectiveness data and recommend improvements.

Metrics data: {aggregated_metrics_json}

Sessions with friction outcomes: {session_ids}

For each underperforming skill:
1. Identify failure patterns from outcome signals
2. Propose specific skill/agent changes
3. Suggest new Iron Laws if patterns are systematic

Write recommendations to: .claude/skill-metrics/recommendations-{date}.md
""")

Step 6: Write Output

Write aggregated metrics to .claude/skill-metrics/dashboard-{date}.json:

{
  "computed_at": "2026-03-03T14:00:00Z",
  "window": "7d",
  "baseline_friction": 0.32,
  "skill_friction": 0.18,
  "friction_delta": -0.14,
  "skills": {
    "/phx:plan": {
      "invocations": 9,
      "trigger_distribution": {
        "user-slash": 9,
        "claude-proactive": 0,
        "nested-skill": 0,
        "unknown": 0
      },
      "proactive_trigger_rate": 0.0,
      "auto_load_gap": true
    }
  },
  "flagged_skills": ["investigate", "plan:auto-load-gap"]
}

Append-only: never modify previous dashboard files.

Iron Laws

1. NEVER modify metrics.jsonl — read-only from this skill
2. Baseline comparison is mandatory — raw numbers without baseline are meaningless
3. Flag, don't judge — surface data, let the human decide what to fix
4. Evidence tags on recommendations — every suggestion needs session citations
5. Trigger source must not be inferred — only treat invocations as user-slash / claude-proactive / nested-skill when the OTel invocation_trigger attribute is present (CC v2.1.126+). Older sessions use "unknown"; never silently bucket them as user-slash — it would hide the auto-load gap.

Integration

/session-scan → metrics.jsonl (with skill_effectiveness)
       ↓
/skill-monitor → dashboard + flagged skills
       ↓
/skill-monitor --improve → recommendations
       ↓
Developer updates skills/agents → deploy → repeat

References

• references/effectiveness-metrics.md — Full metrics schema and evaluation criteria
• references/improvement-template.md — Template for improvement recommendations