PaperWorkflowFig — a Claude Skill

Build journal-submission-grade Fig 1 workflow panels for biology / methods papers. Renders a 4-panel horizontal pipeline (problem → input → novel method → applications) as PDF + JPG via matplotlib, then enforces a mandatory 3-round adversarial review loop. Each round, the JPG is sent in parallel to both an external CLI agent (e.g. Codex) and a Claude subagent; their critiques are merged and you revise.

Why this skill

The Fig 1 of a paper is the first thing a reviewer sees — wrong layout / colour / number-discipline choices can sink a submission. This skill bakes in the six rules that separate a draft from a publication figure, ships a working matplotlib template with placeholders you customise per-paper, and mandates a dual-reviewer loop (Codex + Claude subagent in parallel) that catches roughly 30–50 % more issues per round than a single reviewer.

Install

As a local Claude Code skill (auto-trigger)

# from ~/.claude/skills/
git clone https://github.com/YikaiDong-git/PaperWorkflowFig-skill.git paper-workflow-fig

Then restart Claude Code. The skill auto-triggers on phrases like "make a Fig 1", "workflow figure", "pipeline figure", "schematic", "graphical abstract", "method overview".

As a reference repo (manual paste)

Clone anywhere, read SKILL.md, copy scripts/fig1_template.py into your project's figures/ directory and customise.

Quick start

# 1. Copy the template into your paper project
cp scripts/fig1_template.py /path/to/paper/figures/fig1.py

# 2. Edit the per-paper section at the top:
#    - FIG_SIZE_MM
#    - WIDTH_RATIOS
#    - CLAIMS dict      (every number that appears in the figure)
#    - PANEL_A / B / C / D content
#    - ACCENT (panel c only)

# 3. Render
python figures/fig1.py
# writes fig1.pdf (vector, primary) + fig1.jpg (RGB-composited on white)

# 4. Dispatch the JPG to TWO reviewers in parallel (same turn):
#      a) external CLI agent with -i fig1.jpg
#      b) a Claude subagent (Explore or general-purpose) with the JPG path
#    Each reviewer gets the prompt in references/codex_review_prompt.md.

# 5. Merge the critiques, side with the harsher reading when they disagree,
#    fix all S1 issues, most S2, defer S3.
# 6. Re-render and repeat 2-3 rounds.

What's in the box

paper-workflow-fig/
├── SKILL.md                          # the skill itself (Claude reads this)
├── README.md                         # this file
├── scripts/
│   └── fig1_template.py              # parameterised 4-panel matplotlib template
└── references/
    ├── preflight_checklist.md        # 10-min planning checklist
    ├── postrender_checklist.md       # post-render visual + numerical discipline
    ├── caption_template.md           # 6-sentence-per-panel caption skeleton
    ├── codex_review_prompt.md        # the dual-reviewer adversarial prompt
    └── lessons.md                    # condensed lessons / bugs / discipline

The six rules (the heart of the skill)

1. Horizontal pipeline, weighted panel widths. Novel step gets the biggest panel (~35 %).
2. One accent colour, reserved for the novel step. Default: Okabe-Ito vermillion in panel c only.
3. Equation / matrix glyph is the centerpiece, boxed in accent, with one-line italic caption.
4. Concrete numbers everywhere, with denominators. No "many" / "high" / "strong" without a number.
5. Matrix glyphs with labelled dimensions (not abstract X / y).
6. Render PDF + composite-aware JPG. Visually inspect the JPG every round.

The mandatory dual-reviewer loop

This is the single most important part of the skill — and the one most often skipped.

render -> JPG -> {Codex, subagent} (parallel) -> merge -> revise -> render -> ...

Why two reviewers? They have different blind spots. An external CLI agent has no prior context on the paper and won't anchor on what you've already justified across earlier turns. A Claude subagent can cross-check the figure against the manuscript / methods text if relevant. Running them in parallel costs the same wall-clock time as one. Merge their critiques and side with the harsher reading on disagreements.

Each round catches 5–10 specific issues. Three rounds is usually the right number; a fourth is diminishing returns — ship and gather real reviewer feedback.

Constraints baked in

• No PNG, ever. RGBA → JPG via convert("RGB") leaks colour onto unpredictable backgrounds; the template renders PDF + composite-aware JPG (paste-with-alpha-mask onto explicit white). Also, PNG via some network mounts has corrupted-cache risk; JPG with explicit white compositing is more portable.
• Raw strings for LaTeX math. r"..." everywhere; "\t" becomes a TAB in non-raw strings and ruins \times.
• mathtext subset only. No \tfrac, \dfrac, \bigl, \bigr, \boldsymbol. Stick to \frac, \mathbf, \mathrm, \mathbb, \hat, \Delta, \left( ... \right).
• Render where the file lives. If your project sits on a network-mounted filesystem with write-back caching, render server-side and read back the server-side output path; otherwise local edit + remote render can race.

License

MIT.

Author

Yikai Dong