@ggsvelte/skill

The ggsvelte agent skill: SKILL.md plus deep-dive references/ that teach a coding agent the ggsvelte grammar of graphics — Svelte 5 child-component composition, PortableSpec JSON authoring, and the validate() / ggsvelte-render feedback loop.

This package is the one published home of the skill. It versions in lock-step with @ggsvelte/spec, @ggsvelte/core, @ggsvelte/svelte, and @ggsvelte/cli: a given version number describes the spec, the renderers, and this skill as of the same release. Pin it like any other dependency and let dependabot (or npm-check-updates) tell you when the bundled skill in your repo is stale.

Install

bun add -D @ggsvelte/skill
# or: npm install --save-dev @ggsvelte/skill

The package root is the skill directory: SKILL.md sits next to this README. Skill loaders key off the frontmatter (name: ggsvelte), not the directory name.

Use

Point your agent at the skill, or copy/symlink it into your agent's skills directory under the name ggsvelte:

# Claude Code
cp -R node_modules/@ggsvelte/skill .claude/skills/ggsvelte

# pi
cp -R node_modules/@ggsvelte/skill .pi/agent/skills/ggsvelte

# or reference it in place
node_modules/@ggsvelte/skill/SKILL.md

Re-run the copy on every version bump (a two-line postinstall or a sync script works); the dependabot PR is the signal that the skill changed.

The skill assumes the agent can also run @ggsvelte/cli (ggsvelte-render) for spec validation and headless SVG rendering — install it in every sandbox where an agent authors specs.

Contents

• SKILL.md — trigger conditions, layer ontology, authoring workflow.
• references/ — geoms and stats, scales and palettes, themes, interactions, composition surfaces, recipes.

Guarantees

The skill cannot quietly fall behind the library. CI enforces three layers:

1. Content contracts (scripts/skill-content.test.ts, scripts/skill-package.test.ts) — inventory completeness for every geom, stat, position, theme, and color scheme; every complete JSON fence normalizes and validates; pack shape and lock-step version.
2. Trigger / disclosure contracts (scripts/skill-trigger.test.ts) — frontmatter description quality (the loader's selection signal), balanced positive/negative trigger fixtures under evals/trigger-cases.json, relative-link integrity, and progressive disclosure of every references/ file from SKILL.md.
3. Held-out NL→spec evals (tests/evals/, bun run evals) — model capability on PortableSpec authoring with deterministic graders. These do not load this skill today; agent-in-the-loop skill evals (with/without skill A/B, multi-trial trigger accuracy) are tracked as follow-up work.

evals/trigger-cases.json is not packed to npm (files is only SKILL.md + references/). It is the living seed for future skill-loaded agent evals.