{
  "skill_name": "pmax",
  "source": "aggregated from skills/pmax/evaluations/*.json",
  "evals": [
    {
      "id": 1,
      "name": "askuserquestion_usage",
      "scenario": "/pmax build a prompt for generating weekly status reports from a bug tracker",
      "prompt": "/pmax build a prompt for generating weekly status reports from a bug tracker",
      "files": [],
      "expected_behavior": [
        "Every clarifying question goes through AskUserQuestion. No free-text questions appear in chat.",
        "Chat text between rounds contains only ledger entries and sample artifacts. No explanations, no context, no commentary.",
        "The skill never issues an AskUserQuestion with more than 4 options (schema cap).",
        "The skill never places context, background, or explanation in chat text before or after an AskUserQuestion; all such content lives inside the question body and option descriptions.",
        "Sample artifacts (rendered XML, example prompt output) may appear in chat because the sample is the subject of a decision."
      ]
    },
    {
      "id": 2,
      "name": "db_schema_migration",
      "scenario": "/pmax build a prompt for migrating the theme_db schema so new entries add a theme_db_id primary key and content_id becom",
      "prompt": "/pmax build a prompt for migrating the theme_db schema so new entries add a theme_db_id primary key and content_id becomes nullable, with a backfill for existing rows",
      "files": [
        "Y:/Projects/theme_db/schema.sql"
      ],
      "expected_behavior": [
        "Phase 0 silently reads schema.sql and identifies tables affected before any AskUserQuestion fires.",
        "Phase 2 asks about schema changes in hierarchical rounds where a parent decision spawns child decisions (e.g. chosen primary key type spawns follow-up about backfill strategy).",
        "Every AskUserQuestion option uses plain language; any technical term includes a plain-language explanation in the description.",
        "The final XML contains an <!-- E#.# --> HTML comment before every sentence, with no sentence missing a backing entry.",
        "No XML section appears in the final prompt unless at least one ledger entry populated it.",
        "The Outcome preview bullets in Phase 5 mention every ledger entry present at end of Phase 4."
      ]
    },
    {
      "id": 3,
      "name": "plain_language",
      "scenario": "/pmax build a prompt for an agent that manages NULLABLE foreign keys with ON DELETE CASCADE behavior",
      "prompt": "/pmax build a prompt for an agent that manages NULLABLE foreign keys with ON DELETE CASCADE behavior",
      "files": [],
      "expected_behavior": [
        "Every AskUserQuestion question text uses plain words a non-expert can read.",
        "Every AskUserQuestion option label uses plain words; technical terms in labels are reserved for cases where no plain alternative exists.",
        "Every option whose label contains a technical term includes a plain-language explanation in the description.",
        "The skill never uses hedging words (plausibly, maybe, probably, could) in its own questions or descriptions.",
        "Validator fails and returns to Phase 2 if any AskUserQuestion field contains a banned jargon word without plain-language accompaniment."
      ]
    },
    {
      "id": 4,
      "name": "preview_completeness",
      "scenario": "/pmax build a prompt for generating release notes from a list of merged pull requests",
      "prompt": "/pmax build a prompt for generating release notes from a list of merged pull requests",
      "files": [],
      "expected_behavior": [
        "Every ledger entry that exists at the end of Phase 4 is represented in at least one Phase 5 preview bullet.",
        "The four preview bullets (What it does, Key inputs, Done when, Quick sample) stay under 20 lines total.",
        "The Quick sample bullet shows real executor output, not a description of what the output would be.",
        "If preview coverage fails (an entry is silently dropped), the skill reports which entries are missing and returns to Phase 4 for re-assembly.",
        "The Ship option is always listed first in the Phase 5 AskUserQuestion with the label (Recommended)."
      ]
    },
    {
      "id": 5,
      "name": "scope_discipline",
      "scenario": "/pmax build a prompt for adding a new API endpoint that returns user preferences",
      "prompt": "/pmax build a prompt for adding a new API endpoint that returns user preferences",
      "files": [],
      "expected_behavior": [
        "The skill does not add workflow steps the user did not approve (for example, file-confirmation steps that were not elicited).",
        "The skill states external integrations correctly (outbound handoff direction, validator script invocation path, reuse references).",
        "The skill refuses to emit the final XML if any XML section lacks a backing ledger entry.",
        "Before issuing a new AskUserQuestion, the skill checks the existing ledger to confirm the answer is not already recorded.",
        "If the skill tries to add a section, step, or tag with no ledger backing, Phase 3 validation fails and the skill returns to Phase 2."
      ]
    },
    {
      "id": 6,
      "name": "tiny_rename",
      "scenario": "/pmax prompt for renaming calc_total to compute_cart_total across the codebase",
      "prompt": "/pmax prompt for renaming calc_total to compute_cart_total across the codebase",
      "files": [],
      "expected_behavior": [
        "Skill asks 3 or fewer questions total across all Phase 2 rounds.",
        "Skill does not spawn sub-questions on unrelated dimensions (no tone, no style, no persona questions for a rename task).",
        "The final XML is under 30 lines.",
        "Phase 2 ends automatically once required sections have entries; the skill does not keep asking after coverage is detected.",
        "Every sentence in the final XML has a backing <!-- E#.# --> comment."
      ]
    },
    {
      "id": 7,
      "name": "writing_mechanics",
      "scenario": "/pmax build a prompt for a code reviewer agent that catches DRY violations",
      "prompt": "/pmax build a prompt for a code reviewer agent that catches DRY violations",
      "files": [],
      "expected_behavior": [
        "Every AskUserQuestion question text is written in third person; the skill never addresses the user as 'you' in the question body.",
        "Every option description is written in third person.",
        "The skill never uses hedging words (plausibly, maybe, probably, could) in its own output.",
        "The skill never issues an AskUserQuestion with fewer than 2 options or more than 4 options.",
        "The skill never uses context-dependent shorthand (for example, 'Option B') in permanent artifacts such as the ledger, Outcome digest, or final XML."
      ]
    }
  ]
}