Skip to content

CONTRIBUTING.md

Latest commit

 

History

History
34 lines (26 loc) · 2.1 KB

CONTRIBUTING.md

File metadata and controls

34 lines (26 loc) · 2.1 KB
Raw

Contributing

Keep every contribution focused on one independently installable skill or one collection-level concern.

Add a skill

  1. Choose a unique kebab-case name.
  2. Create skills/<name>/SKILL.md with valid frontmatter whose name matches the directory.
  3. Add runtime references, scripts, templates, or assets only when the skill needs them during use.
  4. Create the corresponding evals/<name>/ directory.
  5. Add positive, negative, and boundary trigger prompts.
  6. Add synthetic, behavior-preserving fixtures when the skill edits code or artifacts.
  7. Run ./scripts/check_skill.sh <name>.
  8. Run ./scripts/run_fixture_checks.sh <name> when fixtures provide executable checks.
  9. Dry-run with ./scripts/install_skill.sh --dry-run <name>, then install with ./scripts/install_skill.sh <name>.
  10. Update the catalogue in README.md and the unreleased section in CHANGELOG.md.
  11. Create coherent local commits with path-specific staging and reviewed staged diffs.
  12. Review all tracked content for secrets, personal paths, private research, and publication rights before pushing or publishing.

Run ./scripts/check_all_skills.sh and ./scripts/run_fixture_checks.sh before sharing a collection-level change. State exactly which optional checks were unavailable.

Change discipline

  • Preserve existing skills and their public behavior.
  • Keep runtime content out of evals/ and collection infrastructure out of skills/<name>/.
  • Do not commit generated evaluation runs.
  • Do not claim trigger accuracy, behavior preservation, or effectiveness without recording the relevant evaluation.

Evaluation gates for skill changes

  • Add a focused synthetic regression whenever an observed failure leads to a runtime change.
  • Rerun the directly affected fixture, trigger, or boundary gates after every runtime modification.
  • Before a release, rerun collection validation, the skill's full fixture suite, and any real-repository gate whose accepted runtime snapshot changed.
  • Record portable summaries of release evidence, but keep raw sessions, temporary repositories, and generated logs in ignored result directories.