{
  "skill_name": "backlog-triage",
  "evals": [
    {
      "id": 1,
      "prompt": "Here's the Northwind checkout backlog exported from Jira: examples/northwind-backlog.csv. Refinement is Thursday. Find the duplicates, tell me what isn't ready, and give me an order for what we pull next. Anything untouched for 30 days counts as stale.",
      "assertions": [
        "Runs the triage script (or states the counts from the export) rather than estimating counts",
        "Each duplicate names both keys, why they are the same work, and which one survives",
        "Rejects or questions candidates that are similar in title but different work, rather than closing every flagged pair",
        "Each item in the proposed ordering has a reason and says whether it is ready to pull",
        "Oversized items are handed to user-story-splitter rather than scheduled as is",
        "Priority changes are presented as proposals with the facts behind them, not applied silently",
        "Ends with a change list of key, change and reason",
        "Does not invent owners, dates, estimates or customer counts not in the export"
      ]
    },
    {
      "id": 2,
      "prompt": "Acme's platform team has 340 open GitHub issues. I ran gh issue list --json number,title,body,state,assignees,labels,updatedAt > issues.json. We don't use estimates or a priority field, just labels like p1/p2. Clean it up.",
      "assertions": [
        "Notes that the estimate and priority columns are missing and that the oversized and priority checks are skipped or mapped from labels",
        "Uses --map or label-based priority rather than guessing priorities",
        "Reports staleness from updatedAt with exact day counts",
        "Confirms duplicates by reading descriptions, not title similarity alone",
        "Does not invent a priority or estimate for items that have none"
      ]
    }
  ]
}