Engineering Research Figure Skill

This skill is a figure advisor for engineering research. It thinks from the
paper claim backward: identify the scientific argument, inspect the data and
experiment design, choose the figure type, generate publication-grade plots,
then audit whether the figure actually supports the claim.

It covers data-driven figures only:

• Raw engineering data and sensor/measurement analysis
• Neural-network training and evaluation results
• Model comparison, ablation, sensitivity, robustness, and efficiency tradeoffs
• Multi-panel paper figures and appendix figures

It does not create mechanism diagrams, architecture cartoons, or flowcharts
unless the user explicitly asks for a separate diagram workflow.

Workflow

1. Clarify the claim. Ask or infer: what should the reader believe after
   seeing this figure? State the assumption before plotting.
2. Classify the task. Use one or more modes:
   • data-analysis: distribution, outliers, correlation, feature behavior
   • training-dynamics: loss, accuracy, learning rate, stability, overfit
   • model-evaluation: regression/classification metrics and diagnostics
   • engineering-comparison: methods, working conditions, parameters, devices
   • ablation-sensitivity: controlled variable changes and robustness
   • paper-figure-plan: multi-panel figure layout and caption mapping
3. Profile the inputs. Inspect files before choosing charts. For tabular
   data, run scripts/profile_experiment.py. For training logs, run
   scripts/parse_training_logs.py when applicable.
4. Choose the figure. Read only the relevant reference:
   • references/chart_selection_engineering.md
   • references/neural_network_figures.md
   • references/experiment_result_figures.md
5. Create a figure plan. For every figure, record claim, data source,
   chart type, axes, grouping, statistics, uncertainty, target venue size,
   export formats, and caption notes.
6. Plot at final size. Prefer matplotlib/seaborn for publication figures.
   Use plotly only for exploratory interactive views or when requested.
7. Audit the result. Check scientific honesty and visual quality using
   references/figure_quality_checklist.md. Render a PNG preview and inspect
   text, legends, overlap, color, axis scales, and panel alignment.
8. Export and summarize. Save vector-first outputs (PDF/SVG) plus PNG
   preview when useful. Include a concise caption draft and any caveats.

Required Reasoning Before Plotting

Never start by drawing the chart the user names if the data design makes it
misleading. Provide a recommended chart, a short reason, and at most two
alternatives. If the user insists, comply only after recording the risk.

Always check:

• What is the experimental unit? sample, run, seed, device, subject, trial?
• Are there repeated runs or independent replicates?
• Are train/validation/test splits clearly separated?
• Are compared methods evaluated on the same x-axis, metric, and data split?
• Is uncertainty needed: SD, SEM, CI, IQR, or seed-to-seed variability?
• Is the figure for the main paper or appendix?

Engineering Plot Guardrails

Actively flag these issues before producing final figures:

• One best training curve shown without seed variability.
• Training results shown without validation/test behavior.
• Mean-only bars for small sample counts.
• Ablations where more than one factor changes at once.
• Methods compared with inconsistent metrics, splits, or x-axis budgets.
• Predicted-vs-true plots without the y = x reference line.
• Confusion matrices without count/normalized labeling.
• ROC/PR curves reported without AUC/AP and class-balance context.
• t-SNE/UMAP used as strong proof instead of qualitative support.
• Too many methods/conditions squeezed into a single unreadable panel.
• Log scale used without noting it in labels or caption.

Figure Plan Template

Use this compact artifact in chat or write it to figure_plan.md when the
project has multiple figures:

## Figure N: [working title]
- Claim:
- Data source:
- Experiment unit and n:
- Mode:
- Recommended chart:
- Alternatives rejected:
- X/Y axes and units:
- Grouping/color/marker:
- Statistics and uncertainty:
- Target venue/size:
- Export:
- Caption notes:
- Risks/caveats:
- QA result:

References

• Use references/caption_and_claim_mapping.md before writing captions or when
  the user is preparing a paper section.
• Use references/attribution.md when describing inspiration, citation, or
  acknowledgements for this skill.

Output Expectations

For analysis-only requests, return a figure recommendation table and the
reasoning. For creation requests, produce the figure files, a figure plan, a
caption draft, and a QA summary. Keep final chat summaries short and state any
limitations clearly.