Proposal: commit-craft skill — commit-series hygiene (split, secret-gate, why-messages), complements changelog-generator

What I'd like to propose

A skill called commit-craft — turns a tangled working tree into a clean, reviewable commit series. Given a mixed diff, it groups changes into logical commits (feature / refactor / test / chore), scans each group for leaked secrets before anything is staged, and writes messages that explain the why, not the what. It's the skill I reach for at the end of every work session, because agents are excellent at producing large diffs and terrible at leaving them in a state a human reviewer wants to read.

Overlap analysis against the existing catalog (this is the part I did first):

• changelog-generator is downstream of commits (turns a merged series into release notes) — complementary, different input moment
• code-reviewer / pr-review-expert / adversarial-reviewer fire after the series is pushed — complementary, commit-craft runs before
• I searched the catalog for a commit-hygiene skill (splitting, staging order, secret pre-scan, message discipline) and found none — the closest artifacts are handoff (session summaries, not VCS) and prompt-governance (different axis entirely)

So: no duplicate, one clear upstream/downstream boundary with changelog-generator, and a distinct trigger moment ("I'm about to commit this mess").

The skill already exists and is MIT-licensed — I wrote it as part of a 15-skill engineering-workflow pack (https://github.com/Hahaknight/claude-skills-pro, 7 of the 15 are free/MIT including this one). Happy to adapt it to your repo's structure and conventions, or to adjust scope if you see overlap I missed. Full text of the current version: skills/commit-craft/SKILL.md in that repo.

Two design choices worth flagging for your review process:

1. The secret scan is a hard gate, not advice — the workflow refuses to stage a group containing credential-shaped strings and requires explicit acknowledgment to proceed. (Your #1001 scan report is exactly the failure mode this is designed to catch at commit time instead of audit time.)
2. Anti-patterns section explicitly bans "one giant commit with a long message" and "commits that split a logical change so finely that each is unreviewable alone" — the two failure modes of over- and under-splitting. Telling the agent what not to do is half the skill.