A Claude Code skill-building framework.

Claude's default knowledge is stack-agnostic, but most developers live inside one stack at a time. The same idea — reactive state, lifecycle, error boundaries, circuit composition — lands differently in React, Vue, Nuxt, and Qiskit, and a generic answer costs round-trips. Collapse exists to close that gap: collapsed skills carry your cross-stack vocabulary so Claude reaches for the right idiom on the first try, with trigger phrases derived from your own annotations.

There's a second, quieter payoff. Writing the Vue version of a lesson after the React version forces you to see where the languages actually diverge — ref is pull-based, mutates .value in place, the wrapper itself is the dependency edge — distinctions you only feel by writing both side-by-side. The lesson captures that, and the compiled skill preserves it.

Three pluggable ingestors feed one typed pipeline. The MDX ingestor reads annotated lessons — code fences with line-level annotation metadata linked to sibling notes, scoped per-stack. The notebook ingestor accepts pasted or uploaded Jupyter .ipynb and MyST markdown, infers the kernel language, and auto-prefills annotations from MyST admonitions. And any new source format ships in about four files following a documented extension pattern.

The template engine compiles each pattern into a SKILL.md: it composes Claude trigger phrases from your annotations and lesson titles, populates cross-language equivalents automatically from sibling stack-tabs, and renders YAML-frontmatter markdown. A three-tier quality linter (clean / info / warn) grades every draft on description length, trigger-phrase ambiguity, and naming — surfaced as verdict dots so weak skills are visible before they ship.

Persistence is local and atomic: temp-file-plus-rename writes into ~/.claude/skills/, path traversal rejected, collisions returned as 409s with the existing description for diff context. No telemetry, no cloud, no database — the filesystem is the storage layer, which is exactly right for artifacts whose whole purpose is to live in your local Claude configuration.

Next on the roadmap: the same pipeline, second output target — MCP server scaffolds. The ingestor layer doesn't change; the work is a template engine that emits a tool surface instead of a skill.