SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillsfigure-qa

figure-qa

Writing & research

View repository

Updated 10-16 06:50

About this skill

Figure QA

Routes a scientific figure to an independent, fresh-context QA reviewer that detects the input type, runs the deterministic checks (fonts, palette, geometry, alpha, resolution, rendered text) with the helper scripts, and adds a vision-language model (VLM) rubric judgment for the aesthetic dimensions. This skill is a thin dispatcher: the detection logic, exit-code contract, VLM rubric, report shape, and the iterate loop all live in references/, and the deterministic engine lives in the figures plugin's agents/figure-qa-scripts/.

When to use

Activate when a figure needs a journal-submission QA pass, or after a figures skill (figures:scientific-figure, figures:svg-figure, figures:svg-primitives, figures:transparent-icons, figures:ai-full-figure, figures:plot-styling) produces output, unless the caller passes no-qa.

What to pass to the reviewer

Every dispatch carries the same briefing, so the right checks run:

  • the figure path and, when known, the input type
  • the target journal (nature, science, cell, pnas, poster, slide, or generic)
  • the theme path (figures/theme.json) so palette compliance is judged against the project bible rather than a fixed allow-list
  • every verbatim string the figure was asked to render (titles, panel letters, labels); without this list the raster text check is silently off
  • the physical width in millimetres for raster inputs, so text height converts to points

Why a fresh-context reviewer

A QA pass is more trustworthy from a reviewer that did not just author the figure. Run it in a separate context and pass only the briefing above. On tools without subagents, run the procedure inline. For several candidates, issue one dispatch per candidate in a single message so they run in parallel on the Sonnet tier.

Dispatch

In every branch the reviewer follows references/figure-qa-procedure.md (strict separation: scripts own ground-truth measurements, the VLM owns aesthetic judgment).

  • Claude Code: Agent(subagent_type: "figures:figure-qa", ...) with the briefing above. Honor a no-qa opt-out by returning immediately.
  • Codex CLI: plugin installation exposes this skill, not a Codex subagent. To use a fresh-context Codex reviewer, first copy ${CLAUDE_PLUGIN_ROOT}/agents/templates/figure-qa.toml (the plugin's agents/templates/ directory) to ~/.codex/agents/ or .codex/agents/, then invoke that configured agent if the current Codex surface supports /agent. If no Codex subagent is configured or available, use the fallback branch.
  • Copilot CLI: plugin installation exposes this skill and, through .github/plugin/plugin.json, the .agent.md reviewer in agents/templates/. Invoke that configured agent when the current Copilot surface supports custom agents. If running outside a plugin install, copy agents/templates/figure-qa.agent.md to .github/agents/ or ~/.copilot/agents/. If no custom agent is available, use the fallback branch.
  • Fallback (no subagent support, or an interactive in-thread check): first locate the procedure ($CLAUDE_PLUGIN_ROOT/skills/figure-qa/references, else find . -type d -path '*/skills/figure-qa/references' | head -1); if it cannot be found, stop and tell the user to install the figures plugin rather than guessing checks. Then follow references/figure-qa-procedure.md directly.

The report and the loop

The report keeps its markdown sections and ends with one fenced JSON block: status (ship, revise, block), findings[] each with check, severity, message, action (regenerate, edit, overlay, rescale, recolor, none) and a hint, plus measurements and the five VLM scores. Generation skills branch on that block. references/iterate-loop.md defines the generate, QA, fix, regenerate loop (N candidates in parallel, one targeted change per iteration, stop at ship or after three iterations) that figures:ai-full-figure follows.

The brain (do not duplicate into dispatch or agent shells)

  • references/figure-qa-procedure.md: no-qa opt-out, script location, type detection, per-branch checks including the raster text check and theme palette compliance, exit-code contract, VLM rubric, the exact report shape, and the finding-to-action table.
  • references/iterate-loop.md: the candidate, QA, fix loop with its worker briefing and stopping conditions.
  • agents/figure-qa-scripts/check_{svg,raster,plot_script}.py (in the figures plugin): the deterministic engine, kept in place so other skills can call it directly. check_raster.py --json --expect-text "..." --palette figures/theme.json --width-mm 89 --journal nature is the text and palette check; check_svg.py --json --palette figures/theme.json covers SVG geometry, fonts, and palette.