Skip to content

feat(claude): add derive-commit-taxonomy skill for designing a repo's commit taxonomy - #782

Open
ppat wants to merge 2 commits into
mainfrom
skill/derive-commit-taxonomy
Open

feat(claude): add derive-commit-taxonomy skill for designing a repo's commit taxonomy#782
ppat wants to merge 2 commits into
mainfrom
skill/derive-commit-taxonomy

Conversation

@ppat

@ppat ppat commented Aug 20, 2026

Copy link
Copy Markdown
Owner

What

Adds private_dot_claude/skills/derive-commit-taxonomy/SKILL.md — a skill that guides an LLM through deriving a commit type/scope taxonomy for a repository it has not seen, plus the enforcement, rule document, and verification that make the taxonomy real.

Why a method, not an answer

No two repos will look the same (especially for scopes): of the two repos this method has been applied to (github-workflows, homelab-ops-kubernetes-apps), the type enums converged but the scope enums share almost nothing. The skill therefore deliberately carries no default enum — it teaches:

  • The questions, in order: what routes a release; what reads each header field and does anything act on it; whether the gate is actually enforced on what lands; what the bot can emit and whether the gate would accept it; which scopes the release layer knows about; what a shared-preset bump does to emission.
  • The traps (the part that took two adversarial review rounds to find): enforcement decomposes per actor and per merge path; a check the dominant actor cannot satisfy is not a check; equivalence classes leak repo state; verify the verifier's own scope; and the calver lesson — an adversarial pass can manufacture a defect by judging a rule against a prior that does not govern the artifact.
  • A three-layer verification shape (closure by construction / runtime gate + declared sampling / falsification with injected defects), including why a raw cross-product and a purely analytic config reading each failed on their own.
  • How to write the resulting rule document so a future session can decide from the diff alone.

The skill states its sample size (two repos) and ends with an amendment protocol: extend it after each repo it is applied to, per the intent that it improves iteratively as overriding patterns emerge.

🤖 Generated with Claude Code

ppat and others added 2 commits August 20, 2026 22:05
… commit taxonomy

A method skill, not an answer: the ordered questions that determine a
repo's commit type/scope taxonomy (what routes a release, what reads each
header field, whether the gate is enforced on what lands, what the bot can
emit, which scopes the release layer knows), the traps that took
adversarial review rounds to find (per-actor merge paths, checks the
dominant actor cannot satisfy, equivalence classes leaking repo state,
verifying the verifier, judging rules against priors that do not govern
the artifact), and a three-layer verification shape. Derived from two
repos; the skill says so and carries an amendment protocol for the next.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…reator methodology

- description: add audit/diagnosis trigger phrasings (Q3/Q4/Q6 standing alone)
- verification: pair each layer with the attack that keeps it honest, moving
  the two verification-time traps to their point of use
- traps: add 'write the intent onto the fence' as the structural prevention
  for the wrong-prior trap, incl. the assertion-vs-workaround danger class
- opening: state verification's claim honestly (closure vs declared sampling)

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant