plugin-review

name	plugin-review
description	Use to analyze plugin usage patterns and suggest improvements. Triggers on "플러그인 리뷰", "plugin review", "사용 패턴 분석".

Plugin Review — Self-Improvement Analysis

Process

1. Collect Data

Run the bundled parsing script to aggregate usage across ALL projects:

python3 {skill_base_dir}/scripts/parse-transcripts.py --all-projects --days 30

This scans every ~/.claude/projects/*/ directory, parses JSONL transcripts, and outputs a single JSON summary to stdout. Stderr shows progress.

The script automatically extracts:

• Skill invocations — both Skill tool calls and /slash-command usage, deduplicated
• Agent dispatches — agent type + model for each Agent tool call
• Token usage — input/output/cache tokens, split by main vs sub-agent, by model
• Skill chains — 2-gram sequences showing common skill workflows
• Per-project breakdown — which projects use which skills/agents

2. Analyze Patterns

Delegate to architect agent with the JSON output. The architect should analyze:

Skill Usage

• Frequency ranking — which skills are most/least used
• Unused skills (0 invocations) — candidates for removal or trigger adjustment
• Slash command vs tool call ratio — indicates whether users invoke directly or the LLM routes
• Most common skill chains (e.g., brainstorming → plan → execute)

Agent & Model Distribution

• Agent dispatch counts by type
• Model distribution: haiku% / sonnet% / opus%
• Category routing accuracy: are executor/explore/archivist routed to appropriate model tiers?
• Naming consistency: detect agmo: prefix vs bare name dispatches

Token Usage

• Total tokens (input + output + cache)
• Main session vs sub-agent token split
• Cache hit ratio: cache_read / (input + cache_read)
• Model-wise cost distribution
• Outlier sessions consuming disproportionate tokens

Per-Project Breakdown

• Session count per project
• Top skills and agents per project
• Projects with plugin installed but barely used

3. Generate Report

Structure the report as:

## Plugin Review — {start_date} ~ {end_date}

### Usage Summary
- Sessions: N (across M projects)
- Top skills: skill1 (X), skill2 (Y), skill3 (Z)
- Agent distribution: executor (A%), explore (B%), architect (C%)
- Model distribution: haiku (X%), sonnet (Y%), opus (Z%)

### Per-Project Breakdown
| Project | Sessions | Top Skills | Top Agents |
|---------|----------|-----------|------------|
| {name}  | N        | skill1    | executor   |

### Skill Analysis
- Full frequency table with slash/tool breakdown
- Unused skills list
- Chain patterns

### Recommendations

#### Skills
- "{skill}" used 0 times → adjust trigger or remove
- "{skill}" triggered N times but cancelled M times → trigger too sensitive

#### Routing
- executor:opus used N times → review if sonnet suffices
- explore:sonnet used N times → most explores are quick, default haiku

#### Tokens
- Average session cost: ~N tokens
- Cache hit ratio: Z%
- Agent token overhead: N% of total

#### Wisdom
- {N} wisdom entries → prune if >50

4. Save Review

Invoke agmo:save-note to save as:

vault/shared/plugin-reviews/[Review] {YYYY-MM-DD} Plugin Review.md

5. Suggest Skill Updates

If specific skills need modification:

• Describe the proposed change clearly
• Ask user for approval before making modifications