| name | outcomes |
| description | Verification-first workflow for rubric-based grading. Use when reviewing a PR, auditing a Supabase migration, validating a Maps/Grounding diff, or QA-ing event/ticket flows against the rubrics in `.claude/outcomes/`. Enforces evidence-over-claims. NOT for writing product code, deploying, or wiring the Managed Agents API.
|
| paths | [".claude/outcomes/**",".claude/docs/best-practices/**","supabase/migrations/**","supabase/schemas/**","src/**/*Map*","src/**/*Grounding*","src/**/*Ticket*","src/**/*Scanner*","tasks/claude-code/**"] |
outcomes — verification-first grading
Purpose
Outcomes is the independent verifier half of the writer/verifier loop.
A writer produces an artifact (PR diff, migration, map UI, scanner flow); a
verifier grades it against a rubric without trusting the writer's reasoning.
The grader's only job is to find evidence.
Phase 1 today is manual: a human or Claude Code reads a rubric from
.claude/outcomes/, runs the verification commands, pastes output, and produces
a per-criterion verdict. Phase 2 wires this to the Managed Agents API
(define-outcomes) — out of scope for this skill.
When to use
- Grading a PR against
.claude/outcomes/pr-review.md before merge.
- Auditing a new
supabase/migrations/*.sql against supabase-migration.md.
- Validating a Maps/Grounding diff (
src/**/*Map*, places client) against maps-grounding.md.
- QA-ing the Events + Tickets flow (Phase 1 gate items) against
events-ticketing.md.
- Any check where the user types "verify this", "grade this", "is this satisfied", "Definition of Done", or "ready to merge".
When NOT to use
- Writing product code →
mde-task-lifecycle, mdeai-executor.
- Deploying →
/deploy-check, /ship.
- Authoring a new rubric from scratch → see
references/rubric-selection.md §"Writing new rubrics" and the meta-rubric in tasks/claude-code/01-outcomes-plan.md §9.
- Auto-running rubrics on a schedule → deferred (Phase 5 of
tasks/claude-code/01-outcomes-plan.md).
- Generic brainstorming, planning, or research.
Verification culture
Evidence priority (high → low):
- command output with exit code line
- tests —
vitest run summary, Test Files N passed
- SQL results —
pg_policies, relrowsecurity, row counts
- screenshots at the documented viewport (375×812 or 1440×900)
- logs / network traces — HTTP status, request IDs
- code inspection — only as fallback, and only when a binary outcome doesn't apply
Claude must never claim verified / tested / passed / deployed / fixed / production-ready without an evidence marker in the same message. The Stop hook .claude/hooks/stop-attribution-gate.mjs enforces this — see references/evidence-rules.md.
Canonical workflow
1. Writer ships a change.
2. Verifier picks the rubric from .claude/outcomes/<name>.md (see references/rubric-selection.md).
3. Verifier runs every command in the rubric's Evidence checklist; pastes output verbatim.
4. Verifier produces the rubric's Output format: "Pass N/M. <summary>" + bullets per failed criterion.
5. If any criterion fails, the writer revises and resubmits.
6. Loop ends at satisfied (all pass) OR max_iterations.
The verifier should not read the writer's commit messages, PR description, or chat history for evidence — only the artifact (code, migration, screenshot, command output).
Supported rubrics
Live files in .claude/outcomes/:
| Rubric | Trigger | max_iterations |
|---|
pr-review.md | Every PR before merge | 3 |
supabase-migration.md | New supabase/migrations/*.sql | 5 |
maps-grounding.md | Map / grounding / Places diff | 5 |
events-ticketing.md | Ticket checkout, scanner, events tables; Phase 1 gate | 8 |
Future rubrics (drafted in tasks/claude-code/01-outcomes-plan.md §5, not yet shipped as files):
edge-function, frontend-ui, rentals-workflow, sponsor-checkout, mastra-agent, production-deploy, phase1-gate.
Evidence requirements
| Surface | Required evidence |
|---|
| PR | npm run floor summary (lint+typecheck+build+test+verify:edge exit codes), git diff --stat |
| Migration | supabase db reset exit 0 + SELECT relrowsecurity FROM pg_class WHERE relname='<T>' rows + SELECT polname, cmd FROM pg_policies WHERE tablename='<T>' rows |
| Map UI | Screenshot at 1440×900 + 375×812; marker count = data length (window.__MAP_DEBUG__) |
| Tickets | Two-POST idempotency probe + SELECT count(*) FROM tickets WHERE stripe_event_id=$1 = 1; 50-buyer oversell probe with quantity_remaining ≥ 0 |
| Edge fn | npm run verify:edge summary, CORS preflight response, 401/403 on missing JWT |
Full details in references/evidence-rules.md.
Manual outcomes workflow
Today (no API), paste this prompt to Claude Code:
Use the rubric at .claude/outcomes/<name>.md to grade this change.
Run every command in §"Evidence checklist", paste output verbatim, and
format the result as in §"Output format". Do not declare a criterion
satisfied without evidence.
See references/manual-outcomes-workflow.md for Fast mode (single iteration, low-risk PRs) vs Full mode (max_iterations default, anything Phase-1-blocking).
Anti-patterns
The grader must reject any of these:
- "looks good", "should work", "probably fixed" — these are claims, not evidence.
- Code inspection in place of running the test (when running the test is possible).
- Migrations marked satisfied without
pg_policies output.
- Map UI marked satisfied without screenshots at the documented viewports.
- Stripe idempotency marked satisfied by reading the function body.
- Reusing the writer's prior session output as "evidence" — the grader runs commands fresh.
- Auto-deploy on
satisfied — humans always sign off on the merge.
Full list in references/anti-patterns.md.
Progressive disclosure
.claude/outcomes/*.md — rubrics (binary criteria, evidence requirements, output format).
references/*.md — operational deep-dives (evidence rules, rubric selection, manual workflow, anti-patterns).
This SKILL.md — routing only.
Related skills
mde-supabase — RLS / migration / edge function patterns that the migration rubric checks against.mde-maps — Maps platform doc map; the maps-grounding rubric cites referer restrictions and Places quota.mde-stripe — payment / webhook / ticket conventions the ticketing rubric grades.testing — Vitest / Playwright / smoke conventions the PR rubric requires.mde-task-lifecycle — owns the writer side of the loop (plan → research → implement → test → ship).security-auditor — paired agent for the security half of the PR rubric.performance-reviewer — paired agent for the perf half of the PR rubric.
Hook coverage
Two Stop hooks reinforce this skill's culture without overriding human judgement:
.claude/hooks/stop-attribution-gate.mjs — blocks turn-end when the assistant claims success without an evidence marker. Negation-aware..claude/hooks/stop-rls-gate.mjs — blocks turn-end when supabase/migrations/** was touched but the final message contains no pg_policies / CREATE POLICY / supabase db reset marker.
If either hook fires, the verifier has skipped a required check. Don't bypass — add the evidence.
