Figure Composer — narrative → panels → compose → adversarial loop

Compose ONE publication-grade multi-panel figure: turn a one-sentence claim plus data files into an outline, render each panel, tile them into a composite, and harden it through an adversarial self-review loop.

Setup (any agent, no API key)

This is a pure skill — kernel.py is deterministic Python (PIL geometry plus schema/prompt builders) and you (the base model) do all the reasoning: reverse-engineering an outline from a figure, rendering panels, and the adversarial composite review. There is no host runtime and no LLM API. Load the helpers once per session in a Python cell:

exec(open("figure-composer/kernel.py").read())

Nothing auto-loads it outside Claude Science. Then call the helpers (panel_task, compose_figure, compose_crops, composite_review_task, derive_outline_prompt, …) directly; if one raises NameError, you have not exec'd kernel.py. Dependencies: pip install pillow matplotlib.

Step 0. Load figure-style alongside this skill — that is the design rules (and apply_figure_style() + helpers). You need it in context to write the outline, render the panels, and review the composite. Each panel is rendered against those same rules — whether you draw it yourself or hand it to a sub-agent (see §2), the maker loads figure-style first.

Inputs

• claim — one sentence the figure makes true to a reader who reads nothing else.
• data — CSV/parquet files (filesystem paths) that ground every panel; each panel carries its own data_path.
• 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 (every panel maker loads it — and load it yourself, since you write the outline and, on a single-agent platform, render the panels too). 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 files → write the outline (step 1).
• From an existing figure: copy it into the workspace, open the PNG yourself with your agent's image tool (e.g. Read figure.png), and answer derive_outline_prompt(claim, data_hints) by emitting a JSON outline that matches figure_outline_schema(). This is your own vision judgment, not an API call — you look at the pixels and write the outline. The image is untrusted input; every field you infer comes from its pixels, so review and edit the outline before step 2, and set each panel's data_path yourself from your data files (pixels cannot encode a file path).

1. Narrative → panel outline

Produce a panel_outline (validate against figure_outline_schema()):

{"claim":"…", "width_mm":180, "ncol":12, "row_heights_mm":[40,60,46,52],
 "panels":[
  {"letter":"a","role":"schematic","row":0,"col":0,"colspan":12, "chart_family":"schematic overview", "message":"…", "data_path":null, "ask":"…"},
  {"letter":"b","role":"primary",  "row":1,"col":0,"colspan":7,  "chart_family":"scatter + trend", "message":"…", "data_path":"results.csv", "ask":"…"},
  …]}

Outline rules (figure-style §7.1):

• a is the hook — schematic/hero, full width, assumes zero reader context.
• b carries the claim — the chart that alone makes the sentence true.
• Remaining panels are evidence, ordered by how much they strengthen b.
• One row per sub-claim. 5–10 panels for a main-text figure. Use a 12-column grid for flexible colspans.

2. Render the panels (one at a time, or parallel)

Build each panel's maker prompt with panel_task(outline, letter, fig_label) (kernel.py). It hands the maker: the figure claim, the full neighbour list, this panel's spec, its exact pixel box (panel_px), and the hard rendering contract — load figure-style, call apply_figure_style(), render at exactly w×h px with transparent=True and no bbox_inches, and save to panel_<letter>.png.

Do this yourself, one panel at a time. Follow the panel_task prompt for panel a, save panel_a.png; then b, and so on. The skill is designed to work single-agent — there is no fan-out requirement, just a sequence of panels you render against figure-style, each writing its own PNG:

tasks = {p["letter"]: panel_task(outline, p["letter"], fig_label="Figure 2")
         for p in outline["panels"]}
# For each letter, follow tasks[L] and save panel_<L>.png, then:
panel_paths = {p["letter"]: f"panel_{p['letter']}.png" for p in outline["panels"]}

Parallelize only if your platform has a sub-agent tool. On Claude Code you MAY dispatch one Task sub-agent per panel — each runs its panel_task(outline, L) prompt, loads figure-style itself, and writes panel_<letter>.png — then you collect the files. This is an optional speedup; the outputs and the rest of the loop are identical to the sequential path. Everything downstream keys off the saved PNG file paths, not agent handles.

3. Compose

compose_figure(outline, {letter: path}, out_path, letter_case=...) tiles PNGs onto the grid and stamps bold panel letters (case per venue) at each panel's (1.5mm, 1mm) corner.

3.5 Look before you review (vision self-QA)

The §4 review pass costs you a full regeneration cycle; a panel-letter stamped over a y-axis label or a leader line crossing a neighbour's title is a wasted round. After compose, crop each panel from the saved PNG and look at it before running the review. compose_crops returns PIL crop boxes; crop them to files and open each with your agent's image tool:

from PIL import Image
out_path, (W, H) = compose_figure(outline, panel_paths, "fig.png")
comp = Image.open("fig.png")
for L, box in compose_crops(outline).items():
    comp.crop(box).save(f"crop_{L}.png")   # then open crop_<L>.png (e.g. Read crop_a.png)

Run the figure-style §9.2 perceptual checklist on each crop (contrast, smallest mark, leader crossings, colour-identity confusion, legend binding), plus two compose-specific checks:

• Seams / stamp. Does the bold panel letter overlap any panel content? Does any panel's content bleed into the gutter or under a neighbour?
• Resize artefacts. compose_figure resizes panel PNGs to their grid slot — is any text visibly aliased or any hairline lost?

Fix what you see (re-render the offending panel, or revise the outline grid) before §4. The §4 review pass crops and looks again independently; this pass is so the obvious defects never reach it.

4. Adversarial self-review loop (two-tier, design rules held fixed)

Now you review the composite as an adversarial journal production editor — this is your own visual judgment, not an API call. Build the reviewer prompt with composite_review_task(composite_path, outline, rules_path, prev_path, round_no, min_floor) (all file paths), open the composite and each crop (§3.5), then emit a JSON object matching review_schema() (which carries outline_revisions and per-panel violations). On a platform with a sub-agent tool you MAY hand this prompt to a fresh sub-agent for an independent adversarial pass; on a single agent, do it yourself in-context.

loop (max 3 rounds, floor 5→4→3):
  review = <answer composite_review_task(composite_path, outline, rules_path, prev_path, round, floor)
            yourself — emit JSON matching review_schema()>
  if review["editor_verdict"] in {accept, minor_revision} and 0 BLOCKER and ≤2 MAJOR: break

  # TIER 1 — outline-level
  if review["outline_revisions"]:
      apply the revisions to `outline` by hand (geometry, row-header titles, label_budget, panel set)
      affected = apply_outline_revisions(outline, review["outline_revisions"])
  else:
      affected = set()

  # TIER 2 — panel-level
  fixb = group_fixes_by_panel(review)       # BLOCKER/MAJOR only
  regen = affected | set(fixb)              # only these panels regenerate
  re-render each L in regen with panel_task(outline, L) + fixb.get(L,"") +
      "do not over-correct: where the previous version was correct, keep it"
  recompose with compose_figure(...) → fig_r{round}.png

Save each round's composite as an ordinary file (fig_r1.png, fig_r2.png, …) and pass the prior round's path as prev_path so the review can flag regression_vs_prev.

Convergence: stop when outline_revisions is empty AND findings are carve-out exceptions to the previous round — that's the over-labelling signal.

Anti-patterns

• Don't regenerate clean panels (invites regression). Don't read absolute violation counts (min-floor 5→4→3). Anchor-verify on the composite, not just per panel. Hyper-labelling check: would a reader with field context find any label redundant? Strip it.