metric-pack-designer

Metric Pack Designer

Use this skill when the user wants to extend cc-plugin-eval with a local rubric.

Invoke the shipped script through the plugin root, which works in any checkout: node "${CLAUDE_PLUGIN_ROOT}/scripts/cc-plugin-eval.js" <command> <args>. Inline mentions of cc-plugin-eval <command> are shorthand for that same invocation; the bare cc-plugin-eval alias only exists if you ran npm link in the plugin directory.

Workflow

1. Clarify the custom rubric categories and target kinds (skill, plugin, directory).
2. Define the smallest useful checks[] and metrics[] payload that answers the user's actual question.
3. Create a metric-pack manifest plus a script that prints JSON to stdout.
4. Wire the pack into a run with cc-plugin-eval analyze <path> --metric-pack <manifest.json>. Repeat the flag for multiple packs.

Manifest Shape

{
  "name": "team-rubric",
  "version": "1.0.0",
  "supportedTargetKinds": ["skill", "plugin"],
  "command": ["node", "./emit-team-rubric.js"]
}

Runtime Contract

• The manifest is passed to cc-plugin-eval analyze --metric-pack <manifest.json>.
• The command runs from the manifest directory.
• The target path and target kind are appended as CLI arguments.
• The process also receives:
  • CC_PLUGIN_EVAL_TARGET
  • CC_PLUGIN_EVAL_TARGET_KIND
  • CC_PLUGIN_EVAL_METRIC_PACK_MANIFEST

Output Contract

The metric pack must print JSON to stdout in this shape:

{
  "checks": [
    {
      "id": "team-style",
      "category": "custom",
      "severity": "warning",
      "status": "warn",
      "message": "Custom rubric finding",
      "evidence": ["detail"],
      "remediation": ["fix detail"]
    }
  ],
  "metrics": [
    {
      "id": "team-score",
      "category": "custom",
      "value": 4,
      "unit": "points",
      "band": "good"
    }
  ],
  "artifacts": []
}

Design Rules

• Keep IDs stable across runs so cc-plugin-eval compare stays meaningful.
• Emit only checks[], metrics[], and optional artifacts[].
• Do not try to overwrite the core summary or the budget bucket totals.
• Prefer deterministic local signals over subjective text generation.
• Use unique IDs that do not collide with the built-in CC1xx-CC9xx codes.