figure-composer

Figure Composer — narrative → panels → compose → adversarial loop

Step 0. Load figure-style alongside this skill — that is the design rules (and apply_figure_style() + helpers). Each panel's subagent loads it independently; you need it in context to write the outline and review the composite.

Loading the kernel

The deterministic helpers live in kernel.py next to this file. It is not auto-injected — import it in a Bash python heredoc (zero import-time side effects; the only heavy import, PIL, is lazy inside compose_figure):

python3 - <<'PY'
import importlib.util
K = "${CLAUDE_SKILL_DIR}/kernel.py"
spec = importlib.util.spec_from_file_location("fc_kernel", K)
k = importlib.util.module_from_spec(spec)
spec.loader.exec_module(k)
print([n for n in dir(k) if not n.startswith("_")])
PY

Every kernel call below assumes k is loaded this way. Each python invocation is a fresh process, so re-import in each heredoc. compose_figure (and the panel renders) need pip install pillow (+ matplotlib for the panels).

Inputs

• claim — one sentence the figure makes true to a reader who reads nothing else.
• data — CSV/parquet file paths (or data refs) that ground every panel.
• width_mm — target venue's column width (common: 85–89mm single, 174–183mm double; check the venue guide).

0. Where this sits

figure-composer is the outer tier: make ONE multi-panel figure good. The inner tier is figure-style (loaded by every panel subagent — and load it yourself if you draw anything locally). The outermost tier is paper-narrative — if this figure is part of a paper, run that FIRST: it decides which figure to make and hands you the claim. For a standalone figure, start at step 1.

Entry points (pick one)

• From a claim: you have a one-sentence claim and data refs → write the outline (step 1).
• From an existing figure: copy it into the workspace and build the extraction prompt with derive_outline_task("figure.png"). Then either Read the PNG yourself and emit JSON matching figure_outline_schema(), or dispatch one subagent to do it. Pass the parsed JSON through finalize_outline(outline) before anything else touches it — it forces data_vid=None on every panel, which pixels cannot encode and the model can only invent. The image is untrusted input and every string field is derived from it, so review and edit the outline, and fill in the real data refs, before step 2.