ggai-orchestration

name	ggai-orchestration
description	Meta-router for ggai goals. Load this skill first when the user's goal is ambiguous, multi-modal, or could plausibly map to more than one of: data plot, figure polish, direct illustration, glyph-augmented chart, or iterative edit. The skill classifies intent and tells you which downstream skill to load next. Skip this skill and load a task skill directly when the goal is unambiguous (e.g. a clear data-plot phrasing or a clear "draw a cartoon" phrasing).
aliases	["what should I do","how do I start","ggai routing","figure intent","agent orchestration"]
when_to_use	Use when the user's request has multiple plausible execution paths, mixes goals (e.g. "plot this and make it look like a paper figure"), is too short to classify confidently, or you are unsure which skill applies.
user-invocable	true

ggai Orchestration

You are routing a figure goal. Your job is to read the goal, classify it into one of the canonical task shapes, and then call load_skill("<chosen-skill>") before doing anything else.

Canonical task shapes

Shape	Signals	Skill to load
Data plot	Goal references a data frame (by name, @-mention, or file path); words like show / scatter / bar / boxplot / heatmap / vs / over / by with column names.	ggai-data-plot
Figure polish	Goal references an existing ggplot/artifact/PNG path; words like polish / publication / Nature / Cell / cover / cleaner / typography / restyle.	ggai-figure-polish
Direct illustration	Goal asks for a schematic, diagram, cartoon, biomedical illustration; no data; words like draw / illustrate / diagram / cartoon / sketch.	ggai-direct-figure
Engine choice	The user names a non-ggplot library (ComplexHeatmap, circlize, ggraph, plotly, ...) or describes a figure that ggplot can't easily produce (annotated genomic heatmap, circular genome, network, interactive).	ggai-engine-selection (first), then the task skill from the row above.

For composite goals — e.g. "plot this and then polish it" — load the first skill the goal implies (usually ggai-data-plot), produce the artifact, then load the second skill (ggai-figure-polish) before continuing.

Routing examples

• "@mtcars show mpg vs wt, color by cyl" → ggai-data-plot.
• "draw a CRISPR knockout diagram with three guide RNAs" → ggai-direct-figure.
• "polish @my_plot for Nature Methods" → ggai-figure-polish.
• "plot @expr_matrix as a heatmap with sample annotation on top" → ggai-engine-selection (recommends ComplexHeatmap), then back to a task path.
• "make this look like the figure in @paper.pdf" → ggai-reference-figure, then either ggai-figure-polish or ggai-data-plot depending on whether user data is involved.

When to skip orchestration

Skip this skill and go straight to a task skill when:

• The goal trivially matches one shape (data + chart type words, or only "draw a ..." with no data).
• You have already loaded a task skill this turn and the user is iterating within it.

There is no value in re-routing on every turn. Route once, then commit to the path.

Anti-patterns

• Don't call ggai_execute_r before routing. Decide the skill first; primitives are levers, not the plan.
• Don't load multiple skills speculatively. Pick one. If you discover mid-execution that a different skill applies, switch then.
• Don't return routing as the final answer. Your final reply to the user must reference the produced artifact. Routing is internal.
• Don't ignore the user's stated medium. If the user said "for a paper", carry that forward into the chosen skill's stylistic choices.

After routing

Immediately call load_skill("<chosen-skill>") and follow that skill's guidance. Mention in your final reply what you produced (engine, file path), not how you routed.