FigMirror (figmirror)

Transfer the visual style of a top-conference paper figure (NeurIPS / ICML / ICLR / Nature / Science) onto user data via an iterative Drawer / Reviewer loop. Output is a self-contained matplotlib script + PNG + type-42 PDF that matches the reference's STYLE — not its data.

When to use

Trigger when the user provides all three:

• A reference paper-figure screenshot.
• Their own data in any parseable form (pasted table, CSV, TSV, markdown

table, or dirty terminal text).

• An expectation that the output should look like the reference — even via

casual phrasing ("make this chart but with my numbers", "redo this in matplotlib"). This includes 3D references when the reference or data is actually 3D.

Do not trigger on plain matplotlib chart requests with no reference image.

Required inputs

• Reference image (PNG / JPG), cropped or uncropped. It may include margins,

captions, neighboring panels, or page text; Stage 0 preprocesses it.

• User data (any parseable form).
• Optional working directory. Default: <cwd>/figmirror-runs/<run-id>/.

3D Insert Gate

Enable references/three-d-prompting.md only when the user asks for a 3D figure, the reference is visibly 3D, or the parsed data requires a 3D encoding such as x/y/z, surfaces, trajectories, layered profiles, closed objects, 3D small multiples, 3D bars, or plane projections. Do not use this insert to turn an ordinary 2D task into 3D.

Architecture

Three bundled subagents drive the loop; the caller orchestrates from the main thread:

• figure-preprocessor — Preprocessor. Stage 0: preserves the raw upload,

crops away margins/captions/page text/neighboring panels when safe, and writes inputs/reference_clean.png plus a crop check/report.

• figure-illustrator — Drawer. Per iter: reads reference + data + L2

library, produces figure_iter<N>.py, img_iter<N>.png, notes_iter<N>.md, floor_selfcheck_iter<N>.txt. Self-checks the layout floor before returning.

• figure-critic — Reviewer. Per iter: vision-only audit on a

fresh-context view (reference + draft + L2 library + optional 3D insert + prior audit only). Returns ONE strict JSON object per the review schema.

Subagents are stateless across dispatch; iter-to-iter state flows through workdir files.

Prefer subagent_type: figure-preprocessor / figure-illustrator / figure-critic. Fallback path when those names don't resolve: see references/iter-loop-spec.md § "Subagent dispatch fallback".

Workflow

For each run:

1. Stage workdir. Pre-create every directory the loop will write into

(subagents Write into existing dirs only — workspace permission quirk). Stage the uploaded reference image to inputs/reference_raw.png and also to inputs/reference_clean.png as a temporary first-paint copy; stage parsed data to inputs/data.txt, and the L2 library to inputs/aesthetic-library.md; stage the 3D router plus references/three-d/ only when the 3D insert gate is enabled. The router selects exactly one mode file: three-d/style-transfer.md for ordinary user-data figures, or three-d/strict-reproduction.md for reproduction, comparison, or candidate/control replacement. For strict 3D reproduction runs that need quantitative candidate diagnosis, also stage the optional candidate scorer. The top-level Orchestrator owns final selection and must run the selected mode's rendered-image gates before copying any candidate to the final figure.

1. Preprocess reference. Dispatch figure-preprocessor before data-gen,

Drawer, or Reviewer. It writes the clean L1 anchor to inputs/reference_clean.png and records the before/after crop check.

1. Echo data parse to user (Decision-7). Show parsed shape (rows × cols,

columns, NaN cells, sample row); proceed when confirmed, or skip if the user pre-authorized. Either way, persist the echo to data_echo.md.

1. Iterate with the caller-provided max_iters; default to 6 when the

caller gives no explicit limit. If the caller enables auto-until-shipped, ignore max_iters and continue until ship or a real blocker. Each iter:

• Dispatch the Drawer.
• Stage the Reviewer's audit view (reference + new draft + L2 library +

optional 3D insert + prior audit only — NEVER data.txt or drawer notes).

• Dispatch the Reviewer.
• Parse the audit JSON.
• Apply the decision rule:

floor.passed && verdict == "ship" → ship and break; else N == max_iters - 1 and not auto → break (fall through to select-best); else continue.

1. Select-best fallback (only if ship never fires). Pick the

lowest-drift iter among floor.passed && verdict == "close" candidates. Document the choice in selection.md.

1. Write canonical artifacts. Copy the chosen-iter script + PNG to

figure.py / figure.png. Re-render figure.pdf with pdf.fonttype = 42.

1. Surface the result to the user. Render figure.png inline, list

paths to figure.py / figure.pdf, give a 1-2 sentence trajectory summary. Do not show audit JSONs or per-iter scripts unless asked.

The full per-step spec (bash commands for staging, dispatch brief templates, audit JSON parsing snippets, drift calculation, fallback selection) lives in references/iter-loop-spec.md. Read it before running the loop.

Non-negotiables

• The reference is a STYLE anchor, not a layout-number anchor. The

Drawer must NOT copy wspace, hspace, figsize, ylim from the reference's data — those recompute from OUR data's shape.

• inputs/reference_raw.png is the preserved upload; inputs/reference_clean.png

is the Stage-0 crop used for L1 measurement.

• Every visual choice traces to L1 (reference image) or L2

(references/aesthetic-library.md). L3 ("I think it would look better") is banned.

• The Reviewer's audit view contains ONLY reference + draft + L2 library +

optional 3D insert + prior audit. NEVER stage data.txt or drawer notes into it. Vision-only audit preserves reviewer independence.

• Pre-create all directories before dispatching subagents; subagents only

Write into existing dirs.

• Iter N>0 Drawer must edit the prior iter's .py incrementally (copy →

edit copy), not rewrite from scratch. All prior iters' artifacts must remain intact in workdir.

• Final script is self-contained (inline DATA SECTOR) with

matplotlib.rcParams['pdf.fonttype'] = 42.

Bundled resources

• references/aesthetic-library.md — L2 convention library (~900 lines).

Read by both Drawer and Reviewer per iter. Versioned independently of the agent prompts because the library iterates faster.

• references/three-d-prompting.md — conditional 3D L2 insert router. Stage

and pass it only when the reference is visibly 3D or the data requires 3D encoding.

• references/three-d/ — 3D mode files plus routed modules for core gates,

surfaces, marks/panels, strict scorecards, and repair feedback.

• scripts/score_3d_candidates.py — optional 3D strict-reproduction helper for

diagnosing rendered camera/aspect/layout candidates against the L1 reference.

• references/iter-loop-spec.md — full per-step orchestration spec

(staging commands, dispatch briefs, JSON parsing, drift calc, fallback).

Bundled subagents

• figure-preprocessor — Stage-0 reference crop role.
• figure-illustrator — Drawer role.
• figure-critic — Reviewer role.

Source: .claude/agents/figure-{preprocessor,illustrator,critic}.md (project-level) or ~/.claude/agents/figure-{preprocessor,illustrator,critic}.md (user-level after scripts/install_claude_skill.py).