Academic Figure Skill

Academic Figure Skill Hub

Academic Figure Skill generates publication-grade scientific figures for Nature/Cell/Science family journals. Every figure starts from the scientific question, not from a template.

Design Principles

1. One figure, one core message. CNS reviewers skim fast. A figure must convey its main conclusion in 3 seconds. Remove gridlines, borders, and legend entries that dilute the core message. Default to minimal, not maximal.

2. Restrained color > abundant color. Default palettes (matplotlib tab10, ggplot2 hue_pal, Excel colors) signal "not designed" instantly. Use 2-4 semantic main colors + 1 accent. See references/color-palettes.md.

3. Design for print, not screen. Journal column widths are fixed (89 mm single, 183 mm double). Set figure dimensions at creation time — never scale down post-render.

4. Vector first, raster as fallback. Line art, scatter plots, bar charts → PDF/SVG/EPS. Only true raster content (heatmap color blocks, micrographs) should use TIFF/PNG at ≥300 dpi.

Complete Workflow

Follow this closed loop for every request. Never skip steps. Never assume the user's question.

User request received
       │
       ▼
  Step -1: Understand the Task ←── DISPATCH FIRST. If the user gave data but
       │                         didn't say what they want to learn, ASK.
       ▼
  Step 0: Parse Data, Match to Task ←── Data parsing is directed by the
       │                                question, not by a template.
       ▼
  Step 1: Recommend & Justify ←── "To answer your question, you need these
       │                         N panels. Here's why." N depends on the
       ▼                         question, not on a fixed number.
  User Confirms ──→ No → Refine recommendation
       │
       ▼ Yes
  Step 2: Runtime & Environment ←── Detect Python + R availability.
       │                            Install or configure if missing.
       ▼
  Step 3: Style Baseline Injection ←── typography + color + export blocks.
       │                               Copied VERBATIM into every script.
       ▼
  Step 4: Production Asset Scan ←── ls assets/figures/. For EVERY panel in
       │                            the plan, check matching scripts.
       ▼
  Step 5: Generate ←── COPY-FIRST for matching scripts → native run.
       │              No match → cross-type inherit.
       ▼
  Step 5.5: Validate Data ←── Data sanity checks BEFORE rendering.
       │
       ▼
  Step 6: QA Protocol ←── 4-pass QA (AP-0..AP-7, CL-1..CL-7,
       │                  VI-1..VI-6, VV-1..VV-5). Fix → re-render.
       ▼
  Step 7: Deliver ←── Vector PDF master + 300dpi PNG preview
                      + QA report + statistics report

Step -1: Understand the Task — DISPATCH FIRST

Before reading a single data file, understand what the user is trying to learn. A dataset without a question is a spreadsheet, not a figure. The entire visualization plan flows from this question.

Classify the request into one of:

If the user gave data without a question, the default response is a question, not a figure.

Step 0: Classify & Parse

Step 0a — Classify the figure archetype FIRST. Before parsing data, classify the figure into one of four archetypes. This determines layout strategy.

If unsure, default to asymmetric_mixed — it produces the Nature-style look of one dominant panel + subordinate evidence panels. Pure symmetric grids are rare in top-tier journals outside of supplement.

Pass the archetype to compose_figure(archetype=...). The engine auto-detects which panel is the hero and applies asymmetric layout when the archetype is not quantitative_grid or symmetric.

Step 0b — Parse data. Once the archetype is classified, read and analyze the data — but only the aspects relevant to the question.

1. Parse structure: Load the data. Count rows, columns. Classify each column (continuous / categorical / group label / identifier).

DATA INTEGRITY RULE — NON-NEGOTIABLE:

• Use ALL user-provided data. Never downsample, never select "representative" rows, never pick "a few columns to keep it simple", never np.random.choice(df, 12) when the user gave 1200 rows. The user's data IS the ground truth. Reducing it is fabricating evidence.
• If rendering performance is a concern (e.g. scatter plot with >50K points), use rasterized=True or vector-friendly density representations. Never subsample to solve a rendering problem.
• If a production script expects a different data shape than the user provided (e.g. script needs real_value.csv + pred_value.csv but user has Pheno1-Pheno8), do NOT silently adapt by picking two columns. Inform the user: "This script expects paired X/Y data. Your data has N columns. Which two should I use as X and Y?" Wait for confirmation before proceeding.
• If you drop any rows or columns for ANY reason, state exactly how many were dropped, why, and get user acknowledgment in the QA report. Silent data loss is the #2 most serious bug in this skill.

2. Run question-directed analysis: The question determines what to compute.
   • "Does X differ between groups?" → Compute group means, between/within variance, pairwise tests
   • "What drives Y?" → Compute correlation matrix, PCA, feature importance
   • "Are these two variables related?" → Compute correlation, fit line, check residuals
   • "What's the distribution?" → Compute summary stats, test normality, detect outliers
3. Do NOT run generic analysis. If the question is about group differences, don't report the full correlation matrix. If the question is about correlation, don't compute group means for every possible partition. Data parsing serves the question, not the other way around.

Step 1: Recommend Figures & Justify

Based on the question and the data morphology, propose a visualization plan. For each panel, state:

1. What question does this panel answer? — One sentence per panel.
2. Why this figure type? — Justify the choice based on data characteristics, not habit.
3. How many panels? — Determined by the number of distinct questions, not by a template. Two questions → two panels. Six questions → six panels. No filler panels.

The recommendation must follow this structure:

Visualization Plan

Scientific goal: [one sentence — what the user wants to learn]

| Panel | Figure type | Answers | Data basis |
|-------|------------|---------|------------|
| (a)   | [type]     | [question] | [e.g. Pheno1-6 all r>0.9 → heatmap] |
| (b)   | [type]     | [question] | [e.g. PC1=92% → PCA with group coloring] |
| ...   | ...        | ...     | ... |

Total: N panels. Layout: [M×N]. Journal: [target].

Re-evaluate after user feedback. If the user says "also show X" or "don't need Y", update the plan and re-present. Only proceed after explicit confirmation.

Step 2: Runtime & Environment

After the plan is confirmed, before generating any code, verify the execution environment.

1. Detect available runtimes:

python --version 2>/dev/null || python3 --version 2>/dev/null
Rscript --version 2>/dev/null || ls "/c/Program Files/R/"*/bin/Rscript.exe 2>/dev/null

Store as PYTHON_AVAILABLE, R_AVAILABLE, R_PATH.

2. Match runtimes to the plan: For each panel in the confirmed plan:

   • If a production script exists (will be checked in Step 3) → the script's language must have a runtime
   • If no production script exists → generate in Python (default) or R (user preference)

3. Handle missing runtimes:

   • R panel requested but R not found → Tell user: "Panel (c) needs R with ComplexHeatmap. Install R or I'll use Python fallback (lower quality)."
   • Python panel requested but Python not found → Tell user: "Python not found. Install Python 3.9+."
   • Missing R packages → install.packages(...) command provided
   • Missing Python packages → pip install ... command provided

Do not silently fall back. The user decides whether to install the missing runtime or accept lower quality.

Step 3: Style Baseline Injection (ALWAYS FIRST)

Load these three files and copy their "COPY VERBATIM" code blocks into the script — in order, at the very top, before any panel logic:

1. references/typography.md — rcParams/theme block. Fonts, spines, ticks, legend defaults.
2. references/color-palettes.md — PALETTE constants. All color variables.
3. references/export-specs.md — font embedding + save function.

Also read references/journal-specs.md for target dimensions (89mm single / 183mm double).

Step 4: Production Asset Scan (EVERY panel)

Run AFTER plan confirmation, BEFORE code generation. For EVERY panel in the confirmed plan:

1. First, read references/directory-map.md. This table maps user language (Chinese + English) to exact assets/figures/<dir>/ paths. Find the user's description in the "Keywords" column → use the exact directory path. This prevents the #2 recurring bug: "柱状图" matching the wrong bar sub-directory.
2. Verify with ls assets/figures/<matched-dir>/ that the directory exists and has scripts.
3. Check the matching directory for production scripts (.py, .R, .r).
4. For each match, check if the script's language runtime is available (from Step 2).
5. Verify script can work with user data. Read the script. Identify the script's "data entry points" — which variables receive external data (column names, data frames, file paths). Map them to the user's data columns. If the mapping exists but columns differ (e.g. script expects length/number_of_cds but user has Pheno1/Pheno2), mark as "visual adapt" — the script's visual system is preserved, only data mapping changes. Only mark as "incompatible" when the data STRUCTURE fundamentally differs (e.g. script expects paired X/Y CSV files but user has a single wide table).
6. Fallback: Only if directory-map.md has no matching entry, fall back to scanning ls assets/figures/ and matching directory names.

Decision per panel:

Panel type matched in assets/figures/<type>/
    │
    ├── Script exists + runtime available + data STRUCTURE matches → COPY-FIRST.
    │   1. Copy the ENTIRE script file to work_dir/<panel>_production.<ext>
    │   2. Find the data-loading line, replace ONLY the data path
    │   3. Execute: subprocess.run([python/r_bin, script])
    │   4. If FAIL: log reason → downgrade to VISUAL ADAPT with notice
    │   5. If OK: load rendered PNG via Image.open() in composition
    │   Panel quality = PRODUCTION. NEVER write a drawing function for this panel.
    │
    ├── Script exists + data COLUMNS differ but STRUCTURE matches → VISUAL ADAPT (NEW).
    │   The script's visual system (layout, coords, colormap, annotations, inset
    │   positioning, statistical overlays) is PRESERVED. Only the data mapping changes.
    │
    │   VISUAL ADAPT protocol — PRESERVES production script's visual system:
    │
    │   **Step 0: SEMANTIC MATCH CHECK (NEW — run BEFORE column mapping).**
    │   Open the production script's companion PNG. Answer:
    │   - What does this script actually PRODUCE? (scatter with marginal density?
    │     grouped KDE? 2D kernel density contour? ridgeplot? violin?)
    │   - What does the USER want? (from their request: "边际密度图", "核密度图",
    │     "分布图", "density", "KDE", etc.)
    │   - Are they semantically compatible? (1D vs 1D? 2D vs 2D? same number of
    │     data dimensions?)
    │
    │   **If YES → Proceed to Step 1 (column mapping).**
    │   The script produces what the user wants. Column names differ but the
    │   visual system (layout, axes, transformations) can be preserved.
    │
    │   **If NO → STOP. Do not VISUAL ADAPT.**
    │   Column mapping cannot fix a semantic mismatch. A JointGrid 2D scatter
    │   script cannot produce a 1D marginal density plot no matter how you map
    │   columns. Two cases:
    │   - User wants 1D density of multiple variables → cross-type inherit from
    │     Ridge (density fill + bandwidth) or KDE/Density (distribution pattern)
    │   - User wants 2D scatter density → find a semantically matching 2D script
    │
    │   1. Identify the script's data entry points:
    │      - What column names does it reference? (e.g. "length", "number_of_cds")
    │      - What grouping variable? (e.g. "type", "Group")
    │      - What filtering columns? (e.g. "is_drep95")
    │   2. Map to user's data:
    │      - Script column X → User column ____ (choose most semantically similar)
    │      - Script column Y → User column ____
    │      - Script grouping col → User column ____ (or None if user has no groups)
    │   3. PRESENT THE MAPPING TO THE USER for confirmation.
    │   4. **Data range compatibility check (NEW):** Before copying the script, verify
    │      that the user's data values are compatible with the script's mathematical
    │      operations. Specifically check:
    │      - Does the script use log-scale axes? → User data must have NO values ≤ 0
    │      - Does the script normalize to [0,1]? → User data must have finite range
    │      - Does the script use sqrt/division? → User data must be non-negative
    │      - Does the script bin continuous data? → User data must have sufficient range
    │
    │      If ANY check fails, identify which visual elements must be adapted:
    │      - Log scale with negative values → drop log transform, keep linear axes
    │      - Normalization with zero range → skip normalization, use raw values
    │      - Division by zero risk → add epsilon guard
    │      Tell the user which transforms were dropped and why.
    │   5. Copy the script. Modify ONLY:
    │      - Column name references (search-replace old col name → user col name)
    │      - Grouping column mapping (or set group=None if none)
    │      - Mathematical transforms that conflict with data range (from step 4)
    │      - Data path
    │      - NEVER modify: layout structure, color palettes, annotation logic,
    │        figure dimensions, export parameters
    │   6. Execute the adapted script. Verify output PNG renders.
    │   7. If execution fails: log reason → downgrade to PARAM INHERIT with notice
    │   Panel quality = PRODUCTION VISUAL SYSTEM. The layout and aesthetics are
    │   from the production script; only the data source changed.
    │
    ├── Script exists + data STRUCTURE incompatible → CROSS-TYPE INHERIT + INFORM.
    │   Tell user: "[dir] has scripts but fundamentally different data structure.
    │   Using cross-type inheritance. Visual quality will be lower."
    │
    ├── Script exists + runtime NOT available → EXTRACT PARAMETERS.
    │   Read the script's FULL content. Open its companion PNG.
    │   Extract Class A/B/C parameters. Cross-type inherit.
    │   Inform user of the quality tradeoff.
    │
    └── No script exists → CROSS-TYPE INHERIT.
        Borrow from similar figure type in the borrowing table below.

Borrowing table (no matching script):

Requested type     Borrow from          What transfers
─────────────────────────────────────────────────────
RDA                PCA                  scatter point size, alpha, ellipse style
Radar              Bar                  color palette ratio, legend style
KDE / Density      Ridge                alpha fill, line color, bandwidth factor
UpSet              Confusion Matrix     fill color, edge color, font size
Single-cell traj.  PCA                  scatter point size, alpha, colormap
Forest             Bar                  line width, marker size, color logic
Sankey             (use own if exists)  node color palette, flow alpha
Heatmap            CorrHeatmap          colormap, colorbar settings
Correlation Scatter Scatter (basic)     point size, alpha, regression line style

Class A/B/C parameter extraction

Class A — Hard Parameters (copy verbatim): Color hex, alpha, line widths, marker sizes, bandwidth, density scale, spacing ratios.

Class B — Scaling Parameters (preserve ratio): Font sizes → scale proportionally to Academic Figure Skill 7pt base. Figure dimensions → scale to journal column width.

Class C — Logic Parameters (copy behavior): Legend on/off, grid on/off, spine visibility, statistical annotation format, frameon.

Preview PNG — HARD RULE

You MUST open the companion PNG before finalizing parameters. Find the PNG with the same basename as the script. Inspect it. Ask: Do my extracted parameters reproduce this visual output? If uncertain → re-read the script.

Step 5: Generate Code

Asset Confirmation Table — MUST be the first lines of the generated script.

Before any import, before any baseline block, the generated script MUST start with this exact comment block:

# Academic Figure Skill Asset Confirmation (verified against assets/figures/)
# (a) [figure type] → [asset path or "cross-type inherit"] → [native run | param inherit]
# (b) [figure type] → [asset path or "cross-type inherit"] → [native run | param inherit]
# ...
# RULE: "native run" = load pre-rendered PNG via Image.open().ax.imshow().
#       "param inherit" = drawing function below that copies Class A/B/C values.
#       If a panel says "native run" and you write a drawing function, you broke the contract.

COPY-FIRST RULE — this is a MECHANICAL rule, not a judgment call.

For EVERY panel marked "native run" in the Asset Confirmation Table:

1. Copy the production script file to a working file. Use shutil.copy() or shell cp. The filename must be <panel_label>_production.<ext>.
2. Modify ONLY the data path. Open the copied file, find the line that reads data (pd.read_csv(...), read.csv(...), read.delim(...)), and replace the path with the user's data path. Change NOTHING else.
3. Execute the copied script. Python: subprocess.run([sys.executable, script]). R: subprocess.run([r_bin, script]).
4. Verify output. Check that the output PNG/PDF file exists and is non-zero. If execution fails (syntax error, missing package, data mismatch), log the error, set that panel to "param inherit" in a REVISED Asset Confirmation Table comment, and proceed with cross-type inheritance instead. The failure reason MUST appear in the QA report.
5. For R scripts: ensure png(type="cairo") and showtext_auto(FALSE) before png() per references/r-rendering.md.

What this means in the composition script:

• "native run" panels → NO drawing function. The script loads the pre-rendered PNG via Image.open().
• "param inherit" panels → drawing function with Class A/B/C values extracted from the named asset.
• If a native run FAILED → REVISED table shows the downgrade to "param inherit" and the reason.

Forbidden patterns (these indicate the COPY-FIRST rule was violated):

• A drawing function whose name matches a panel marked "native run" in the table.
• A drawing function that is a "simplified version" of a production script.
• Importing functions from a production script instead of executing the whole script.

Script structure:

1. Asset Confirmation Table (MANDATORY — the first lines)
2. Style baseline (verbatim from Step 3) — ONCE at top
3. Data section — ALL user data loaded here, NEVER downsampled
4. Production script execution (for "native run" panels) — subprocess calls
5. Panel functions — PNG loaders for "native run", drawing functions for "param inherit"
6. compose_figure() call — handles layout, spacing, labels, export

Mixed Python+R: R panels run first to PNG at spec-correct dimensions. Python engine loads them via ax.imshow().

R PNG rendering: Follow the three mandatory rules in references/r-rendering.md. Never skip them — they prevent the #1 recurring R PNG quality bug.

Step 5.5: Validate Data Before Rendering

Run chart-type checks on every panel's data array. Fix failures before rendering.

Volcano: (padj < 0.05).sum() ≥ 10 AND ≤ 80%
AUROC:   (tpr - fpr).max() ≥ 0.15
Heatmap: np.std(data, axis=1).mean() > 0.2
Bar:     abs(means.max() - means.min()) / means.max() > 0.05
Corr:    abs(corr_mat).max() (off-diag) ≥ 0.3
PCA/RDA: between_group_var / within_group_var > 1.0
Box/Violin: abs(medians.max() - medians.min()) / data_range > 0.1

Also predict 3 visual problems at the target panel width. Fix them before rendering.

Step 6: QA Protocol

Execute the 4-pass QA protocol from references/checklist.md.

• Pass 0: Anti-pattern scan (AP-0 through AP-7) + revision case cross-reference (VI-7)
• Pass 1: Code-level compliance (CL-1 through CL-7)
• Pass 2: Visual logic & data integrity (VI-1 through VI-6)
• Pass 3: Rendered output verification (VV-1 through VV-5)

QA Gateway: READY (all pass) → Deliver. FIX (≤2) → Fix and re-run. ESCALATE (>2) → Reviewer Simulation Mode.

If Python/R runtime is unavailable, skip Pass 3 and warn the user.

Step 7: Deliver

Output:

1. Complete code block
2. QA self-check report
3. Vector PDF master + 300dpi PNG preview
4. Statistics & Reproducibility Report — for every quantitative panel:
   • n definition (what does each replicate represent?)
   • center statistic (mean? median?)
   • spread / interval (SD? SEM? 95% CI? — define which one was used)
   • statistical test name
   • multiple-comparison correction (if applicable)
   • source-data traceability (which file/column produced each data point)
   • For ML/model figures: train/validation/test split, number of seeds/folds, metric definition, baseline definition
5. Suggested next step

The statistics report is not optional. A figure without statistical documentation is not a publication-grade figure. Reviewers expect this information and will request it if missing.

Reviewer Simulation Mode

If the user asks "will this pass review?":

Role: You are a reviewer for [journal]. Review through 5 lenses:

1. Scientific clarity — main message in 3 seconds?
2. Visual hierarchy — most important element draws the eye?
3. Color & accessibility — print-safe, colorblind-friendly?
4. Typography & labeling — legible at print size?
5. Overall polish — anything making you "frown"?

Output: structured feedback with priority (must-fix vs. suggestion) and fix instructions.

Anti-Pattern Recognition

Flag these immediately when reviewing user code:

• Default color palettes → offer semantic palette
• Four-sided borders → offer clean left/bottom spine
• Legend inside plot → offer outside/direct labeling
• Screenshot-only export → offer vector PDF
• Jet/rainbow colormap → offer perceptually uniform
• Bar chart without individual points (n < 10) → offer strip/swarm overlay
• Default font → offer Arial/Helvetica

Figure Type Routing

All figure types routed through assets/figures/<type>/. Adding a new type = adding a directory with scripts. No registration needed.

Eval & Quality Tracking

# Full asset audit — scripts, syntax, baseline compliance
py academic-figure-skill/scripts/eval_runner.py
py academic-figure-skill/scripts/eval_runner.py --type PCA

# Trigger accuracy benchmark — 40 prompts, F1/precision/recall
py academic-figure-skill/scripts/trigger_benchmark.py
py academic-figure-skill/scripts/trigger_benchmark.py --calibrate

# QA validator coverage — test each check against known-good/bad scripts
py academic-figure-skill/scripts/qa_coverage.py

# Reference integrity — directory-map, PANEL_ASPECT, SKILL.md cross-links
py academic-figure-skill/scripts/check_references.py
py academic-figure-skill/scripts/check_references.py --json

# E2E integration — automated A/B scenario scoring
py academic-figure-skill/scripts/e2e_runner.py
py academic-figure-skill/scripts/e2e_runner.py --scenario S1_pca

Cross-Platform Adapters

Generate platform-specific adapter files for non-Claude-Code agents:

python academic-figure-skill/scripts/generate_adapters.py            # all platforms
python academic-figure-skill/scripts/generate_adapters.py --target cursor  # Cursor only

Generated adapters are in install/:

• install/cursor/.cursorrules → copy to your project root for Cursor
• install/copilot/copilot-instructions.md → copy to .github/ for GitHub Copilot
• install/codex/manifest.yaml + instructions.md → copy to ~/.codex/skills/academic-figure-skill/ for Codex
• install/claude-code/README.md → already supported natively via ~/.claude/skills/

References

Always Load

On-Demand

Production Assets