Contributing to Step Beyond

Thanks for wanting to make agents less literal. Ground rules:

Before changing behavior, read the normative specification in SPEC.md.

What we accept

New domain trees (skills/step-beyond/references/domains.md) — must include all six lines: TRIGGER, RECALL, L1, L2, L3, VERIFY, NEVER, and be backed by real usage, not speculation.
Slop patterns (references/slop.md) — concrete, observed patterns with the "why it reads as slop" reasoning.
Eval cases (evals/cases.md) — every behavioral claim in the spec should eventually have a case that can falsify it.
Framework installation recipes (references/installation.md) — tested on the actual framework, with the exact config that worked.
Translations — keep README_PL.md (and future languages) in sync with README.md; PRs that change one without the other will be asked to sync.

Token discipline. This skill's core promise is behavioral impact per token. Every added line in SKILL.md or core-injection.txt must earn its tokens; deep material goes to references/ (progressive disclosure).
Run the evals. Changes to the core instruction, ceiling, or any protocol require a fresh eval run (see evals/README.md) — at minimum all B and D cases. Attach the results file to the PR.
No slop. Contributions are checked against references/slop.md. Yes, really.
One concept per PR. A new domain tree and a README redesign are two PRs.

Behavioral changes to the core instruction → minor bump (or major if pipeline shape changes).
New reference content, examples, evals → patch.
Update CHANGELOG.md, SKILL.md frontmatter version, and .claude-plugin/plugin.json together.

MIT. By contributing you agree your contribution is MIT-licensed.