gabrielmoreira/agent-skills-mirror/figure-composer

Figure composer

Load figure-style with this skill. The sidecar provides pure geometry, composition, task-building, and review-schema helpers. It does not call models, delegate Agents, resolve artifacts, or inspect images from Python.

Inputs

Require a one-sentence claim, target width in millimetres, and concrete project-relative or absolute data paths. Never use artifact ids as paths. For an existing figure, inspect the real image with view_image and write the outline yourself; pixels cannot reveal the source data path.

Workflow

1. Build an outline matching figure_outline_schema(). Put real paths in

data_path; use null for schematics.

1. Make panel a the conceptual hook and panel b the primary evidence. Use a

12-column grid and one row per sub-claim.

1. Build one instruction per panel with panel_task(...).
2. If delegate_tasks is advertised, submit the independent panel tasks as one

batch. Grant each task the minimum advertised capabilities needed, normally visualization plus project_read. Require a concrete PNG filename in each output schema. If delegation is unavailable, render the panels sequentially with python.

1. Compose returned paths with compose_figure(...). Do not pass placeholder

markers to the composer.

1. Save the compose_crops(...) boxes with figure-style's

save_panel_crops(composite_path, compose_crops(outline)), then call view_image on the composite and every crop. Fix seams, clipped labels, aliases, empty space, and misplaced panel letters before review. The crops are inspection debris, not products: they live in .cache/figure-style/, never beside the composite or under the output figures directory, and get deleted once the composite passes.

1. Build one reviewer instruction with composite_review_task(...). Delegate it

with image_inspection, project_read, and reasoning when those capability ids are advertised; otherwise perform the review in the current Agent.

1. Apply outline revisions and regenerate only affected panels. Stop after three

rounds or when there are no blockers and at most two major findings.

Outline example

{
  "claim": "Treatment restores the disease-associated trajectory.",
  "width_mm": 180,
  "ncol": 12,
  "row_heights_mm": [42, 60],
  "panels": [
    {
      "letter": "a",
      "role": "schematic",
      "row": 0,
      "col": 0,
      "colspan": 12,
      "chart_family": "study schematic",
      "message": "The experiment tests trajectory rescue.",
      "data_path": null,
      "ask": "Show cohorts, treatment, sampling, and comparison."
    },
    {
      "letter": "b",
      "role": "primary",
      "row": 1,
      "col": 0,
      "colspan": 12,
      "chart_family": "trajectory plot",
      "message": "Treatment moves cells toward the healthy trajectory.",
      "data_path": "results/trajectory.csv",
      "ask": "Plot disease, treated, and healthy cells with confidence bands."
    }
  ]
}

Boundaries

• Use delegate_tasks only as an explicit Wisp tool; never call delegation from

python.

• Use view_image only on a concrete local image file.
• Keep data preparation in normal project files. Use run_in_context only when

a deterministic render or preprocessing job is long enough to require a persisted Run; Agent delegation itself is not a Run.

• Save the accepted composite to a stable project path and report that path.