AI Full Figure

Generate single panels or whole multi-panel figures with the image model (gpt-image-2 through the Codex CLI, or the OpenAI Images API when a key exists), with the titles, panel letters, and short labels rendered by the model at large size, then run figures:figure-qa and iterate until the figure ships. The output is an opaque PNG per panel plus a composed SVG, PDF, and PNG when more than one panel is built.

Sister skills are invoked by their fully qualified figures:<skill> names through the host's skill-invocation mechanism. Keep host-specific command or tool syntax in the host adapter, not in shared skill instructions.

The workflow

1. Bible first. A project keeps one figures/theme.json. If it does not exist, invoke figures:figure-bible before generating anything. The theme supplies the palette, typography, style and negative tokens, the text limits, and the Codex model settings.
2. Generate. One panel with generate_figure.py, or a whole figure with build_figure.py and a spec file. Text that the model should render is passed as verbatim items with a role and a placement.
3. QA. Invoke figures:figure-qa on every candidate, passing the expected verbatim strings, the theme path, and the journal. The report ends with a JSON block whose status and findings[].action drive the next step.
4. Iterate. Follow the iterate-loop reference shipped with figures:figure-qa (iterate-loop.md): N candidates in parallel, QA in one parallel dispatch, pick the best, apply one targeted change, re-QA, stop at ship or after three iterations.