Cut Release

SkillDocs & knowledge

Cut a new TMDb release — work out the next SemVer version from the evidence, do the pre-tag housekeeping a tag would otherwise freeze in place, draft release notes, then tag and publish the GitHub release. Presents the version (with its reasoning) and the full notes for approval and STOPS; nothing is tagged or published until you say go.

Available today. Use it from your connected AI after setup.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use the Cut Release skill

What this skill tells your AI

The instructions your AI receives, as published by adamayoung/tmdb in .claude/skills/cut-release/SKILL.md and read by ahel’s review.

Takes main from "a pile of merged PRs" to "a tagged, published release".

This skill is never headless. Every other skill in this repo may decide for itself; this one may not. A tag is the one artifact consumers resolve against, and publishing one is effectively irreversible: SwiftPM hands a new minor or patch to everyone inside the current from: range the moment it exists, a new major reaches everyone who has since bumped, and deleting a tag does not un-fetch it from anyone who already resolved. Phases 1–5 are read-only and local. Phase 6 is a hard stop. Phases 7–9 run only on an explicit go-ahead.

If invoked with no human able to answer, run to the hard stop, print the summary, and exit. Do not tag.

Phase 1 — Preflight

Stop on any failure; report which one and what to do.

  • On main, clean tree, in sync with origin/main (git status --porcelain, git rev-list --count main..origin/main).
  • git fetch --tags; establish the last release: git tag --sort=-v:refname | head -1.
  • CI is green on the current main tip: gh api repos/adamayoung/TMDb/commits/<sha>/check-runs. This is the entry check — the commit that actually gets tagged does not exist yet (Phase 7 creates it), and Phase 7 re-checks that one. Starting from a red main just means finding out later.
  • There is something to release — at least one commit since the last tag that is not purely a tag-housekeeping commit.

Phase 2 — Work out the version

Two independent signals. They must agree, or you stop.

Signal A — the CHANGELOG (authoritative). The open ## [X.Y.Z] section at the top (no date = unreleased) is the proposal. Count **Breaking:** markers inside it, and read the section headings:

  • any **Breaking:**major
  • otherwise any ### Added content → minor
  • otherwise (### Fixed / ### Changed only) → patch

Signal B — the commits (cross-check). git log <lasttag>..HEAD, bucketed by gitmoji. Consumer-visible: feature, 🐛 fix, ♻️ refactor, 🔒 security, ⚡️ perf. Not consumer-visible: 🔧 chore, 📝 docs, 👷 CI, tests, 📦 build. Count multibyte-safely — sort | uniq -c over emoji collapses distinct ones in a non-UTF-8 locale and will hand you a confidently wrong tally.

Gitmoji alone is not enough to pick a version in this repo, and you must not try. A 🐛 here is routinely source-breaking: "🐛 Surface task cancellation as TMDbError.cancelled" (PR #433) was filed as a fix and added an enum case, which breaks every exhaustive switch downstream. Signal B exists to catch omissions from the CHANGELOG, not to compute the bump.

Then reconcile:

  • The proposed section heading must match what Signal A computes from its own content. A section headed [20.0.0] containing no **Breaking:** is a mis-labelled release — stop and report, do not silently downgrade it.
  • knowledge/next-major.md's status line names the open window. If it disagrees with the CHANGELOG heading, stop.
  • The computed tag must not already exist.

Coverage check. Every commit since the last tag that touches Sources/ and carries a consumer-visible gitmoji should be represented somewhere in the open CHANGELOG section. List any that are not. This is the check that catches a real change shipping invisibly; the not-consumer-visible buckets legitimately have no CHANGELOG entry and are not flagged.

Phase 3 — Pre-tag housekeeping

Branch first. git checkout -b chore/prepare-<version>-release. Phase 1 put you on a clean main, and CLAUDE.md forbids editing there — but the practical reason is re-entrancy: if the run is abandoned at the Phase 6 stop, edits made on main leave a dirty tree that fails this skill's own Phase 1 preflight on the next attempt. Branching now costs nothing and makes an abort free.

A tag snapshots the tree, so anything that says "unreleased" is frozen saying it, forever, inside the very release it describes. These four are not optional, and all of them were needed for 19.0.0 (PR #406):

  1. README.md install snippet. from: "<previous major>" resolves to >=prev, <prev+1 — on a major, readers following the README would never resolve to the new version at all. Bump it. (This is the one with real user impact; the other three are hygiene.)

    Scripts/check-readme-version.py enforces this from make lint and CI: the snippet must match either the newest dated CHANGELOG section (the last release) or the newest undated one (the release being prepared). That second arm exists so this housekeeping PR is not blocked by the tag it is preparing — bump the snippet and date the section in the same commit and both arms stay satisfied. Forget the bump entirely and the check goes red on the next PR after the release.

  2. CHANGELOG.md. Date the section — ## [X.Y.Z] - YYYY-MM-DD — and add the link definition at the foot: [X.Y.Z]: https://github.com/adamayoung/TMDb/releases/tag/X.Y.Z. Check the previous version's definition is there too; they have been missed before.

  3. knowledge/next-major.md. Rewrite the status line: this version has shipped, and the next major's window is now open. Remove every backlog entry that shipped in this release (the file's own rule). An entry left behind claiming to await a window that just closed is invisible exactly when it should fire.

  4. ADR statuses. Every knowledge/decisions/*.md reading Accepted (in X.Y.Z, unreleased) becomes Accepted (shipped in X.Y.Z) — in the file and in the decisions/README.md index. Leave unreleased — tooling only ADRs alone; they describe .claude/ machinery with no library impact and no version to ship in.

Make these edits on the branch but do not commit yet — they are shown at the hard stop, and they go up as a PR, never straight to main.

Phase 4 — Draft the release notes

Structure, in this order:

  1. The opener. Two or three sentences, warm and a little funny. Write a new one every time — see Voice below.
  2. Upgrade impact, if this is a major or has any **Breaking:** entry. What breaks, and the one-line fix for each. This sits above the feature list.
  3. What's new / What's fixed / What changed — consumer-facing prose, drawn from the CHANGELOG (which is already written for humans), not from commit subjects.
  4. Commits, inside a <details> block: every commit since the last tag, one line each — <sha> <subject>. Collapsed, because it is a reference, not a read.
  5. Full Changelog compare link: https://github.com/adamayoung/TMDb/compare/<lasttag>...<newtag>.

Voice

Slack's iOS notes are the reference: they open with something human, and they never pretend a bugfix release is a moon landing. Aim for a colleague writing to other developers — dry, specific, occasionally silly, never markety.

  • Good: naming the actual absurdity. "This release is mostly about a movie database telling us, with total confidence, that Bryan Cranston has never been in a TV series."
  • Good: cheerful understatement. "Eighteen of twenty tagged images were quietly going in the bin. They are no longer going in the bin."
  • Avoid: "We're excited to announce", "packed with", "🚀", exclamation marks in rows, and any claim the diff does not support.
  • Avoid: whimsy about the breaking part. Joke about the bug you fixed, not about the work you have just made someone do. If this release breaks builds, the opener stays light but stays short — then get straight to the impact.

Length: the opener is 2–3 sentences. If it needs a fourth, it is a paragraph, and paragraphs are what section 3 is for.

Phase 5 — Assemble the summary

Put together, for one message:

  • Version and the reasoning — the deciding evidence, not a recitation. "Major: 14 **Breaking:** entries, including four public enums gaining cases" beats "there were breaking changes".
  • Anything Signal B flagged as missing from the CHANGELOG.
  • The four housekeeping edits, as a diff summary.
  • The complete release notes, verbatim as they will be published.

Phase 6 — STOP

Present the summary and wait. Do not tag, push, or publish anything.

Ask plainly whether to go ahead. Treat anything short of a clear yes as a no — "looks good" about the notes is not approval to publish. If the answer is a change request, apply it and return to this phase; the gate re-arms every time.

Phase 7 — Land the housekeeping

Only after approval. The branch already exists from Phase 3.

Commit the Phase 3 edits as 🔧 Prepare the <version> release, open the PR via /pr, and let its CI run. The narrowed docs/config gate usually applies here — the diff is Markdown only — but /pr owns that call; do not pre-empt it.

Merge once green.

Then wait for main's own run. ci.yml triggers on push to main with **/*.md among its paths, so the squashed merge commit gets a CI run of its own — a different commit from the PR head, and the one the tag will point at. Phase 1's "green on the exact commit being tagged" means this run, not the PR's. Poll it:

gh api repos/adamayoung/TMDb/commits/$(git rev-parse HEAD)/check-runs \
  --jq '[.check_runs[] | {name, status, conclusion}]'

Red or still running → do not tag. This is the cheapest place in the whole skill to be patient.

Phase 8 — Tag

On the merged main, at the housekeeping commit:

git checkout main && git pull
git tag -a <version> -m "<version>"
git push origin <version>

Annotated, not lightweight — a release tag should carry an author and a date.

Regenerate the commit list before publishing. The notes were drafted in Phase 4, before the housekeeping merge existed — so their <details> list is short by exactly the commit being tagged, while the compare link below it includes it. Re-run the log against the new tag and refresh that block.

Phase 9 — Publish

gh release create <version> --repo adamayoung/TMDb \
  --title "<version>" --notes-file <path> --verify-tag --latest

--verify-tag refuses to invent a tag that does not exist, which is what turns a typo into a failed command instead of a phantom release.

Then confirm: print the release URL, and re-read the published notes once to check nothing was mangled in transit.

After

Report the tag, the release URL, and what the next window is now open for. If any knowledge/next-major.md entries were deliberately carried forward rather than shipped, name them — that is the moment someone can still object.

Failure modes worth knowing

  • A red tag. Tagging a commit whose CI never passed. Phase 1 exists for this; do not skip it because "the PR was green" — the merge commit is a different commit, and it gets its own run.
  • Assuming a bad tag is cheap to undo. Two active tag rulesets cover ~ALL tags with creation, update and deletion rules. The repository-role bypass means the owner can force it, but nothing here is a casual git push --delete — and SwiftPM consumers who already resolved the version are unaffected by any of it. Get it right going in.
  • A frozen "unreleased". Skipping Phase 3 ships ADRs and a next-major.md that insist, inside the release, that the release has not happened.
  • A README nobody can follow. The from: bump is invisible in testing and breaks discovery for every new consumer on a major.
  • Version invented from commits. Deriving the bump from gitmoji rather than the CHANGELOG will under-call a major in this repo, roughly every time.

Signals

GitHub stars
176
Forks
47
Last commit
Aug 2026
Advanced
Catalog kind
skill
Gateway key
cut-release
Source
github.com/adamayoung/tmdb