/figure-bake

This engine bakes a single figure from a one-line concept to a manuscript-grade artifact. It is the 3-agent loop demonstration of the orchestration pattern: each iteration runs design (@figure-descriptor) → implementation (@analysis-implementer, which actually executes Python/matplotlib code via Bash) → critique (@reviewer, which reads the rendered PNG/PDF multimodally), with a verdict at the end of every iter.

If you are reading this because Claude Code's skill auto-discovery surfaced it: invoke it via /figure-bake <fig-id>. Do not edit figures.json, the rendered figure files, or the per-figure script directory by hand while a run is in flight.

When this skill is invoked, immediately execute the workflow below. Do not only restate or summarize these instructions back to the user.

Signature

/figure-bake <fig-id> [--max-iter N] [--data <path>] [--code-dir <path>]

Flag	Default	Purpose
`--max-iter`	`3`	Hard cap. Reaching it without DONE/BLOCKED → HALT. Override via CLI or via `## Research stack` field `figure_bake.max_iter`.
`--data`	(resolved)	Dataset root. Three-layer resolution: CLI `--data` > env `CLAUDE_RESEARCH_DATA_ROOT` > CLAUDE.md `## Research stack` `data_root` > hardcoded `./data/` (Phase 2 decision §3).
`--code-dir`	`null`	Optional script directory the implementer writes its renderer(s) into. If unset, the implementer chooses (typically `<manuscript_root>/figures/code/` or alongside `vector_path`).