critic-looplisted
Install: claude install-skill Zane-0x5a/remotion-director
# critic-loop
This stage turns N rendered draws into one refined piece. Two protocols govern it; both are the **authoritative wording** and are used **verbatim** (`⟨…⟩` = slots the orchestrator fills, the rest unchanged). Do NOT paraphrase them — Read the files and use their exact text.
## 1 · Blind select — pick the most promising base
Read **`${CLAUDE_PLUGIN_ROOT}/skills/critic-loop/BLIND-SELECT-PROTOCOL.md`**. Spawn the `blind-selector` agent (fresh, one-shot) with the brief + the N candidate dirs (each has 6 `still-*.png` + `strip/`). It selects for **potential** (the base whose ceiling after the loop is highest), not fewest current flaws — fixable execution nits must not count against a strong base. It returns `{ winner, reason }`. The orchestrator does NOT judge; it hands over candidates and takes back the winner.
## 2 · Critic loop (甲乙环) — refine the winner until it converges
Read **`${CLAUDE_PLUGIN_ROOT}/skills/critic-loop/CRITIC-PROTOCOL.md`** (条带规格 harness contract + the verbatim 甲 prompt + the per-round ferry message + 乙环纪律).
- **甲 (critic)** = the `aesthetic-critic` agent. Spawn ONE instance for the piece and continue the SAME instance across rounds (it is persistent — retained memory across rounds is the point — but design-blind: it sees only the frame strip + the brief, never DESIGN.md or code). It reports phenomena + severity and never prescribes a fix; its last line is `CONVERGED: YES|NO`.
- **乙 (builder)** = the SAME continuous-context builder instance