Stats
Actions
Tags
From skill-eval
Use LLM-as-judge to analyze friction clusters from M2 classified events and identify which skills, prompts, tools, or workflows are the best candidates for optimization. Clusters friction by root cause, scores targets by frequency x severity x improvability, and outputs a ranked target list.
How this skill is triggered — by the user, by Claude, or both
Slash command
/skill-eval:target-identifyThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
LLM-as-judge analysis to surface ranked optimization targets from session friction clusters.
LLM-as-judge analysis to surface ranked optimization targets from session friction clusters.
# Identify targets from classified events
bun scripts/target-identify.ts -i classified-all.jsonl -o targets.jsonl
# Limit to top 3 targets
bun scripts/target-identify.ts -i classified-all.jsonl --top 3 -o targets.jsonl
# Use a specific model
bun scripts/target-identify.ts -i classified-all.jsonl --model claude-sonnet-4-5-20250514 -o targets.jsonl
# Dry run: show clusters without calling LLM
bun scripts/target-identify.ts -i classified-all.jsonl --dry-run
# Stats only: show cluster sizes
bun scripts/target-identify.ts -i classified-all.jsonl --stats
# Require more evidence per cluster
bun scripts/target-identify.ts -i classified-all.jsonl --min-events 5 -o targets.jsonl
# Pipe from stdin
cat classified-all.jsonl | bun scripts/target-identify.ts --top 5 -o targets.jsonl
| Flag | Description |
|---|---|
--input, -i <file> | Input classified-events.jsonl from M2 (or stdin) |
--output, -o <file> | Output file (default: stdout) |
--model <model> | Anthropic model (default: claude-haiku-4-5-20251001) |
--min-events <n> | Minimum events per cluster to assess (default: 2) |
--top <n> | Output top N targets (default: 10) |
--concurrency <n> | Parallel LLM calls (default: 5) |
--stats | Print summary stats only |
--dry-run | Show clusters without calling LLM |
--help | Show help |
@anthropic-ai/sdk (npm)ANTHROPIC_API_KEY environment variableTargets are ranked by composite score: frequency * log2(session_count + 1) * severity * improvability. The LLM judge assesses each friction cluster for root cause, target type, severity (1-5), improvability (1-5), suggested action, and eval questions.
npx claudepluginhub asragab/asragab-claude-marketplace --plugin skill-evalCreates, edits, and optimizes skills for Claude Code, including drafting, evaluating with test prompts, iterating on performance, and improving skill descriptions for better triggering accuracy.