Contributions welcome — especially new stack translations, sharper values, and counter-examples.
- Principle first, code second. Every reference file teaches the medium-agnostic rule before showing stack-specific code. A PR that adds a framework recipe without the principle it serves will be asked to add it.
- Exact values or it didn't happen. "Use generous spacing" is not teachable; "between-group gap ≥ 2× within-group gap" is. Claims should be checkable.
- Keep SKILL.md files short. They are routers (frontmatter + workflow + a
load table). Depth goes in
references/. If a SKILL.md approaches 100 lines, move content down. - Descriptions are triggers. The frontmatter
descriptionis what makes a skill fire — when editing one, keep both halves: what it does and the phrasings that should trigger it. - No lock-in. Examples land in 2–3 different stacks (web, native/GPU, and a plotting/game context). Adding a stack is great; replacing the others isn't.
Copy (or symlink) skills/* into ~/.claude/skills/, start a fresh Claude Code
session, and confirm the edited skill still triggers on its description's
phrasings and that its references resolve. For substantive changes, run the
package's own acceptance test: build a small deliberately-mediocre artifact, run
the full workflow (direction → implement → render → critique → iterate), and
include the before/after in the PR.
Semver. Reference-file improvements are patches, new references or skills are minor versions, format/structure changes are major.