universe-studio — one tag, four approvals

The user's ideal UX: come with an idea, tag a workspace, make FOUR approval decisions while everything else auto-assembles and auto-repairs against the universe's own eval rubric. You are the orchestrator that delivers that. Your input is an idea + an active workspace; your output is a rendered, eval-cleared deliverable produced through four user-approved stages.

This skill is a sequencer, not a craft skill. It does not know how to write a scenario, prompt an image, or cut a video — the per-role playbooks own that. It owns the ORDER, the per-stage eval gate, the bounded repair loop, and the four user checkpoints. At each stage you delegate the actual work to the matching playbook, then gate the result.

ALSO FIRE

• When a workspace is already active (ralphy workspace use <ws> was run earlier this session) and the user drops an idea expecting the staged flow.
• After ralphy workspace eval lands a clearing scorecard for one stage and the user says "good, next stage" / "keep going" (any language — match their chat language, English on disk).
• When the user explicitly opts into batch repair ("just fix it, don't ask me each stage") — you still announce paid spend before it runs, but the per-stage loop may apply paid items.

DO NOT FIRE

• For a workspace WITHOUT an evaluators.json + stageGates rubric — there are no gates to drive. Route to the normal intake / producer flow instead, and (if it fits the user's intent) suggest authoring a workspace evaluator config first.
• For a single isolated stage ("just rewrite scene 3", "regen this one keyframe") — that is the scenarist / art-director playbook directly, not the four-stage orchestrator.
• For a quick one-off render with no universe gating — that is intake → producer.
• As a replacement for /evaluator or /fixer on an arbitrary mp4 — those are the standalone eval + repair skills. This skill only drives the per-stage workspace-eval gates inside the four-stage flow.

HARD INVARIANTS

These are inherited from AGENTS.md; they are load-bearing here because the whole point is unattended auto-repair between checkpoints.

1. Wait-for-go before paid generation (the paid gate). The #473 loop auto-applies only FREE editor fixes (costEstimate === 0). The moment a stage needs a PAID regeneration (costEstimate > 0 — an art-director re-roll, any ralphy generate …), the loop STOPS and surfaces pendingPaidActions. You present them and wait for explicit user approval before spending. No paid call between checkpoints without a prior "batch repair, don't ask me" opt-in. This mirrors AGENTS.md invariant #1 and /fixer's hard gate.
2. ralphy is the only entry-point. Every assemble / eval / repair / render step routes through a ralphy verb. No raw ffmpeg / curl / bunx tsx. If a stage needs a verb that doesn't exist, propose adding it and stop (AGENTS.md #2).
3. Append-only. Re-assembles and repairs auto-version (.v2); never --force-overwrite without the user asking. Each stage's prior workspace-eval.json is archived to .vN; the old artifacts all stay on disk (AGENTS.md #14).
4. Read MODELS.md before any model call (AGENTS.md #6). Claude's training is stale.
5. Generic — no universe literals. Everything (which stage owns which criteria, the gate severity, the benchmarks) flows from the ACTIVE workspace's rubric. Never bake one universe's facts into the flow.

Where the framework lives

You are the agent layer on top of the framework that already landed:

• #468 — the workspace evaluator config (evaluators.json + stageGates) per workspace.
• #469 — ralphy workspace eval <project> → writes workspace-eval.json + workspace-eval-report.md (append-only). The per-criterion scorecard you read at each stage.
• #470 — the six builtin criteria: material-density, edit-correctness, insta-metric-fit (deterministic), scenario-fidelity, character-design-cohesion, location-consistency (vision).
• #472 — the workspace stageGates map stage → contract phase → owned criteria, surfaced as stage-gate-unmet stop conditions by ralphy project status <id> --contract.
• #473 — runStageRepairLoop (cli/lib/eval/stage-loop.ts): the bounded per-stage assemble → eval → repair → re-eval loop with the free=auto / paid=stop discipline.

The stage→phase→criteria map below is the DEFAULT the rubric ships with. Always read the active workspace's actual stageGates — a universe can wire its own gates. Do not hardcode the table.

The four stages

The per-stage loop (run identically for all four)

For each stage, in order:

1. Assemble. Delegate the craft to the owning playbook — read it fully first (docs/playbooks/{art-director,scenarist,editor}.md; researcher for site-grounded reference work). Produce the stage's artifact through ralphy verbs only.
2. Eval. ralphy workspace eval <id> — it scores the project against the workspace rubric and writes workspace-eval.json (append-only). The criteria[] array carries one { id, verdict, findings[] } per criterion; you care about the ones this stage's gate owns.
3. Read the per-criterion scorecard. Pull the owned criteria from the scorecard. Cross-check with ralphy project status <id> --contract — a stage-gate-unmet stop condition names exactly which owned criterion FAILED (block) or WARNED (advisory) and which phase it gates.
4. Repair (bounded, #473). When an owned criterion isn't clean, run the bounded loop discipline: FREE editor fixes (costEstimate === 0 — loudnorm, recut, caption regen, timing) auto-loop until they clear or the retry budget (default 3) is spent; the moment a fix is PAID (costEstimate > 0 — art-director re-roll, any ralphy generate …), STOP and surface pendingPaidActions. Apply fixes through the owning playbook's verbs (/fixer's routing applies — art-director regen / scenarist rewrite / editor recut). Never iterate the same failing gate more than twice without stopping to report options (AGENTS.md #4).
5. Present — only on a clear (or a blocking decision). When every owned criterion clears, show the user the stage's deliverable + a one-line scorecard summary + what the loop auto-fixed. If the loop hit a paid gate or exhausted its budget, surface the blocking decision instead (the pendingPaidActions / the residual findings) and the worst-case cost — the user decides whether to approve the spend or change direction.
6. Wait for the stage approval. Do NOT advance to the next stage until the user approves this one. This is one of the four decisions the whole skill exists to collect. Log a stage skip, if the user forces one, to user-prompts.jsonl (ralphy project log-prompt --stage "skip:<phase-id>", per the production contract).

After stage 4 clears and the user approves, hand off to Unit formation (ralphy unit / /templater) per the production contract — the studio flow ends at an eval-cleared, user-approved render.

How this sits with the production contract

The four stages are a user-facing VIEW of the canonical production contract (docs/playbooks/agent-production-contract.md): stage 1 ≈ phases reference-gate → style-lock, stage 2 ≈ scenario, stage 3 ≈ assets, stage 4 ≈ render → eval. The contract still owns the full phase order and per-phase artifacts; this skill collapses them into four approval checkpoints and adds the per-stage workspace-eval gate. Use ralphy project status <id> --contract as the source of truth for where the project actually is — never guess the phase from chat memory.

References

• cli/lib/eval/stage-loop.ts — runStageRepairLoop: the bounded loop + the free/paid split (the mechanism behind every stage's repair step).
• cli/lib/contract.ts — CONTRACT_PHASES (the real phase ids) + deriveStageGateStops (the stage-gate-unmet stop conditions).
• cli/lib/eval/workspace-criteria.ts — the six builtin criteria + their validator ids.
• .agents/skills/fixer/SKILL.md — the eval-to-repair routing the per-stage repair step reuses.
• .agents/skills/evaluator/SKILL.md — the standalone eval skill (this one drives the workspace-eval variant per stage).
• docs/playbooks/{art-director,scenarist,editor}.md — the per-role playbooks each stage delegates to.
• docs/playbooks/agent-production-contract.md — the canonical phase order the four stages map onto.
• MODELS.md — read before any model pick.