scoring

name	scoring
description	Use after any singularity-managed skill execution to rate performance 0-100, track quality over time, and trigger repair when scores drop below threshold

Score a Skill Execution

Evaluate skill performance using a structured 5-dimension rubric. Scores drive the evolution loop: low scores trigger repair, high scores enable crystallization.

Workflow

Step 1: Identify the Skill

Determine which singularity-managed skill was just executed. Check ~/.claude/singularity/registry.json to confirm it's tracked.

If the skill isn't registered, ask: "This skill isn't tracked by singularity. Want me to register it first?"

Step 2: Assess Performance

Dispatch the singularity-claude:skill-assessor agent (haiku model, fast and cheap) with:

• Skill name and version
• What was requested (the user's original task)
• What was produced (the skill's output/changes)
• The scoring rubric from references/scoring-rubric.md

The assessor returns a structured JSON score.

Step 3: Record the Score

"${CLAUDE_PLUGIN_ROOT}/scripts/score-manager.sh" add <skill-name> <total-score> \
  --context "<what the skill was used for>" \
  --strengths '["<strength1>", "<strength2>"]' \
  --weaknesses '["<weakness1>"]' \
  --edge-cases '["<edge-case-if-any>"]'

Step 4: Check Thresholds

Read config from ~/.claude/singularity/config.json:

Condition	Action
Average < autoRepairThreshold (50) for 2+ runs	Suggest: "This skill is underperforming. Run /singularity-repair to fix it."
Average >= crystallizationThreshold (90) with 5+ runs	Suggest: "This skill is ready for crystallization. Run /singularity-crystallize to lock it."

Step 5: Update Registry

Update ~/.claude/singularity/registry.json with:

• lastExecuted: current timestamp
• executionCount: increment
• averageScore: from score file

Step 6: Log Telemetry

"${CLAUDE_PLUGIN_ROOT}/scripts/telemetry-writer.sh" log <skill-name> \
  --trigger "scoring" \
  --score <total-score> \
  --summary "<brief assessment>"

Step 7: Report

Show the user:

Score: <total>/100 (avg: <average>/100 over <count> runs)
  Correctness:    <n>/20
  Completeness:   <n>/20
  Edge Cases:     <n>/20
  Efficiency:     <n>/20
  Reusability:    <n>/20
Maturity: <level> → <new-level-if-changed>

Scoring Modes

From config.json:

• "auto" — Dispatch assessor agent automatically (default)
• "manual" — Ask user for the 0-100 score directly
• "hybrid" — Auto-assess, show result, let user override