changelog-from-commits

Changelog from Commits (Live)

Users don't read commit messages, and they shouldn't have to. In Claude Cowork this skill reads the real commit range and turns it into a changelog written for the people who use the software — grouped, deduped, and translated from "refactor: extract helper" into what actually changed for them.

What This Skill Produces

• The changelog — grouped into Added / Changed / Fixed / Breaking, each entry written as user-facing benefit, not commit-speak
• The breaking-change callout — migrations and removals pulled to the top with what to do
• A changelog artifact — Keep-a-Changelog style, ready to paste into CHANGELOG.md or a release

Required Inputs

Ask for these if not provided:

• The range — from tag/ref to tag/ref (default: last tag → HEAD), and the repo
• The audience — end users, API consumers, or developers — translation depth follows
• Version & date — the release number and date for the heading

Framework: Commits → Changelog

1. Group by impact — Added / Changed / Fixed / Breaking; drop internal-only churn (chore/ci/refactor) unless it changes behaviour.
2. Translate — every entry says what the user can now do or no longer suffers, not the implementation.
3. Dedupe & merge — many commits behind one feature become one line.
4. Breaking first — removals, renames, and migrations lead, with the upgrade step.
5. Credit & links — PR numbers/authors where it helps.

Execution (Cowork)

1. Read the range — via the GitHub connector, list commits (and merged PRs) between the two refs, with messages, PR titles, and labels.
2. Classify — map each to Added/Changed/Fixed/Breaking; set aside pure-internal churn; detect breaking changes from !/BREAKING CHANGE/removed-public-API signals.
3. Translate & merge — collapse the commits behind each user-facing change into one benefit-led line; keep PR references.
4. Order — breaking first, then Added, Changed, Fixed.
5. Emit the artifact — the changelog block; offer to prepend it to CHANGELOG.md or attach to a release, only on request.

Guardrails: include only changes actually present in the range — never invent a feature; base "breaking" on real signals, not guesses; don't overstate impact; if the connector is unauthorised, work from a pasted git log and say the range couldn't be read live.

Output Format

A Changelog block:

## [version] — date

### ⚠ Breaking
- what changed → what to do (#PR)

### Added
- user-facing capability (#PR)

### Changed
- what's different now (#PR)

### Fixed
- the problem that's gone (#PR)

Quality Checks

• Every entry maps to a real commit/PR in the range
• Entries read as user benefit, not commit messages
• Breaking changes lead and include the migration step
• Internal-only churn was excluded (unless it changed behaviour)
• Multiple commits behind one feature are a single line

Anti-Patterns

• Pasting raw commit messages as the changelog.
• Inventing a feature not in the range to round it out.
• Burying a breaking change in the middle of "Changed".
• Listing every chore:/refactor: that users never see.

Example Trigger Phrases

• "Write the changelog since the last tag in Cowork."
• "Draft release notes from my commits for v2.0."
• "What changed for users between v1.4 and HEAD?"
• "Summarise this commit range as a human changelog."