Menú
How to invoke a rubric and emit structured rubric scores. Loaded by judge agents. The 13 standard-aligned rubrics live in the daemon registry; mirror copies under `.claude/rubrics/<id>.md` are generated via `pp-daemon dump-rubrics`.
Instalar con Codex o Claude Copia este prompt, pégalo en Codex, Claude u otro asistente, y deja que revise la página de la skill y la instale por ti.
Tiered cross-vendor vs same-vendor judge policy. Read this before invoking any judge so the gate is enforced correctly. Loaded on demand by the `pair-programmer` skill. Authoritative source is `mcp__pp_harness__gate_eligible_judges`; this document explains why the policy is the way it is.
Master skill for the pair-programmer harness. Loaded by every /pp:* slash command. Defines the full Phase-11 request lifecycle — triage, profile snapshot, taxonomy mapping, judge routing with cross-vendor policy, Reflexion ×1, missability gate, master-plan patching, and run finalization. Read this before driving any harness run.
How `<project>/.harness/profile.yaml` modifies gates. Loaded by the `profile-loader` agent and referenced by the master skill. On first run, the harness detects the project type and writes profile.yaml after a one-line confirmation; only an explicit `skip` (or pre-existing absent file the user opts not to bootstrap) leaves the run in generic mode.
File layout under `<project>/.harness/<run_id>/` and the kinds of artifacts produced by each stage. Loaded on demand by the `pair-programmer` skill. Every file written under .harness/ MUST go through `mcp__pp_harness__archive_artifact` UNLESS it lives inside an active best-of-N candidate worktree (those files are tracked via git inside the worktree, not via the artifacts table).
Skill for producing distinctive, professional-grade game-design artifacts (GDD, mechanic specs, level greybox, narrative bibles, economy spreadsheets) that avoid generic AI aesthetics and standard "explore a world / engage in combat / build relationships" pitfalls. Invoked by spec-author, narrative-designer, level-designer, encounter-designer, economy-designer when the active profile is in the game-dev family.
Protocol for the `master-plan-patcher` and `run-finalizer` agents. The per-project `<project>/PROJECT_MASTER.md` follows the Section 9 20-section template. Every successful run patches the relevant sections with cross-references to the run's artifacts.
Basado en la clasificación ocupacional SOC
| name | rubric-application |
| description | How to invoke a rubric and emit structured rubric scores. Loaded by judge agents. The 13 standard-aligned rubrics live in the daemon registry; mirror copies under `.claude/rubrics/<id>.md` are generated via `pp-daemon dump-rubrics`. |
Every judge verdict pins a rubric_id. The driver gets the right rubric id from mcp__pp_harness__gate_eligible_judges (field rubric_id); the judge agent then fetches the rubric body via mcp__pp_harness__get_rubric(id) and applies it.
Rubric ids are version-suffixed: wcag-2.2-aa@1, owasp-asvs-l1@1, c4-system-context@1, etc. Pinning the version means a verdict is reproducible against the exact rubric it was scored with.
// request
{ "id": "wcag-2.2-aa@1" }
// response
{
"id": "wcag-2.2-aa@1",
"kind": "design",
"version": "1",
"title": "WCAG 2.2 Level AA",
"source_url": "https://www.w3.org/WAI/standards-guidelines/wcag/",
"markdown": "# WCAG 2.2 AA rubric\n\nScore 0..1 for each principle...",
"schema_json": { /* JSON Schema for the score object */ }
}
If get_rubric returns null, the rubric id is wrong or the rubric was deleted from the registry. Fall back to the closest match (list_rubrics returns all 13 with their kinds) and surface a warning.
The judge agent passes the rubric markdown into the cross-vendor or same-vendor critique tool as the rubric_md parameter. The Codex/Gemini critique helpers force a structured response:
{
"outcome": "pass" | "fail" | "revise",
"critique_md": "...",
"score": { "<dimension-1>": 0.0..1.0, "<dimension-2>": 0.0..1.0, ... }
}
The dimensions are the bullets at the top of every rubric body (e.g. perceivable, operable, understandable, robust for WCAG). The judge MUST score every dimension named in the rubric — verdict-rubric-coverage (PostToolUse hook) warns if fewer than 3 dimensions are present.
Then call mcp__pp_harness__record_verdict:
{
"attempt_id": "attempt_xxx",
"judge_producer": "gemini",
"judge_model_id": "gemini-2.5-pro",
"rubric_id": "wcag-2.2-aa@1",
"outcome": "revise",
"critique_md": "Missing focus-visible state; states 4/8 named.",
"score_json": { "perceivable": 0.8, "operable": 0.55, "understandable": 0.7, "robust": 0.7 }
}
The daemon computes cross_vendor from judge_producer vs attempt.producer. The driver does not pass that flag.
[0.5, 0.7) and no rubric-specific must-have failed (this is the band where Reflexion ×1 is most likely to help).< 0.5, OR any rubric-specific must-have absent (each rubric body names which ones).Rubric-specific must-haves examples:
< 6/8 states named in the screen-state matrix.schema_validity < 0.7 (an invalid spec is not a valid contract).{system_boundary, users_personas, external_systems} < 0.5.musts_clear < 0.5 (vague requirements aren't requirements).definition or grain < 0.5.roles_responsibilities or incident_response < 0.5.eval_suite_present or hitl_thresholds < 0.5.The verdicts table stores rubric_id per verdict. Replay (pp-daemon replay <run_id>) reconstructs the run including the exact rubric body that was used. If a rubric is updated to v2, prior runs are unaffected — they pin v1.
You can drop a markdown file at .claude/rubrics/<custom-id>.md and reference it as a rubric_id; the daemon will fall back to disk if the registry has no match. Keep the dimension list explicit so judges score consistently.