Agent Skills: Artifact Guidelines

Artifact Guidelines

This skill defines the form of artifacts subagents produce. For what to write, see the relevant methodology skill. For which folder, see orchestration > Canonical Structure.

Format choice

• Final reports. Any phase deliverable — EDA, critique, posterior predictive, population assessment, final synthesis — uses HTML. See references/html-report.md.
• Logs and intermediate notes. log.md, README notes, working scratch — use Markdown. See references/markdown-report.md.
• Never use plain .txt.

Figure conventions

• Use descriptive filenames (group_washout_curves.png, not fig1.png).
• One concept per figure; multi-panel only when the comparison demands it (max 2×2).
• Save next to the report that cites them; reference via relative path in <img src>.
• 300 DPI for report figures, 150 DPI for exploratory plots.

File minimalism

Generate fewer, better files:

• One Stan model per .stan file.
• One logical step per Python script (compile, fit, diagnose, plot), not one monolithic analysis. See python-environment > Script Structure.
• Use descriptive script names (fit_hierarchical_model.py, not model.py).
• One consolidated report per phase, not ten partial analyses.
• Combine related visualisations into multi-panel figures when comparison is the point.
• Remove intermediate scratch after consolidating insights into the report.

Every file you create should justify its existence: will this be read? Does it convey unique information?