| name | changelog |
| description | Maintain FreeCut's user-facing weekly changelog manually. Use when the user asks to add a changelog entry, draft bullets from recent commits, or close out a week into a release. Curates commits into polished user-facing bullets and keeps CHANGELOG.md and src/data/changelog.json in sync. |
Changelog skill
FreeCut's changelog is maintained manually — there is no auto-append CI. The operator (or this skill, on request) curates commits into user-facing bullets and updates both files together.
Two synchronized files:
CHANGELOG.md — human-facing, Keep-a-Changelog style, read on GitHubsrc/data/changelog.json — typed, imported by the "What's New" dialog UI
Always update both or neither.
Structure: weekly releases + one rolling current entry
Two tiers:
current — rolling entry for the in-progress week. Shown in the UI as a "This Week" card. Updated manually whenever the user asks to refresh it from recent commits.releases — one entry per completed week, newest first. Each has a version and a date.
Versioning: CalVer, Monday-start weeks
- Format:
YYYY.MM.DD where the date is the Monday that opens the week (Mon–Sun).
- Rollup is manually triggered, typically on Monday morning.
- Mid-week hotfix needing its own release: add
.N suffix (2026.04.13.2). Rare.
package.json version is pinned to 0.0.0 and ignored — the package is private, nothing reads it. The week version lives only in changelog.json and CHANGELOG.md.
No git tags. The web app ships continuously via Vercel.
Separate packages (future CLI/API) get their own semver and their own changelog under their package directory. This file is for the web app only.
Preconditions (run before any mode)
Three checks block all other work. All must pass before drafting, appending, or rolling up.
1. Current must not span past weeks
current is the in-progress week — Mon–Sun, anchored to this week's Monday. If current.date falls in an earlier ISO week than today, you are rollup-overdue: the bullets in current belong to one or more closed weeks, not to the present.
Compute this week's Monday from today (in repo-local time): monday = today - ((today.getDay() + 6) % 7). If current.date is not in [monday, monday+6], stop and run rollup mode first, even if the user asked for append. Do not append into a stale current — it merges this week's bullets with last week's and corrupts both.
If current spans multiple closed weeks (the bullets were appended across several Mondays without rollup), do partitioning rollup. Do not merely re-split the bullets already in current — those bullets are only what someone happened to append, and a skipped week means they are incomplete (this is how the 2026-06-15 transcription and Color-workspace launches were nearly lost: they landed after the last append and were absent from current entirely). Instead, re-derive each week from git: walk git log <branch> --no-merges per Monday–Sunday window from the last release forward, curate each week's commits independently, then reconcile that against the existing current bullets (the existing bullets are a hint, not the source of truth). Emit one [YYYY.MM.DD] release per closed week and start a fresh current for this week.
2. CHANGELOG.md heading must match current.date
The markdown header ## [Current] — week of YYYY-MM-DD is a literal string and does not auto-derive from JSON. Any time you write current (append, rollup, partition), recompute the heading as ## [Current] — week of <Monday-of-current.date> and update it. Never leave the heading frozen at an old Monday — that is the bug that caused the 2026-04-13→2026-05-02 drift.
3. current + releases must fully cover the git log since the last release
The skill's default mode is incremental append, which trusts current to already hold everything prior and only adds newly-mentioned commits. That trust breaks silently when a week (or an append) is skipped: those commits never enter the changelog and nothing flags them — the outcome then depends on how often someone remembers to run the skill. Close that gap by reconciling against git on every run, so infrequent triggering is lossless instead of lossy.
Before drafting/appending/rolling up, walk git log <branch> --no-merges --since=<date-of-last-release> and confirm every user-facing commit (after curation rules) is represented in either current or an existing release. Surface anything missing to the user and fold it in — do not assume current is complete just because it has bullets. This is the check that would have caught the missing 2026-06-15 launches at append time instead of three weeks later. The reconciliation is against the curated git log, not a raw commit count: dropped noise (chore/ci/test/refactor, same-week regressions, follow-ups) is expected to be absent and is not a gap.
Modes
The user invokes one of these explicitly. Default to asking which mode if unclear.
Draft mode (most common)
The user asks "draft changelog bullets for this week" or similar. Produce bullets from commits — do not write files yet, just show the user the curated list to approve.
- Pick the branch to read from:
- For the current rolling week (drafting bullets for
current), query the active development branch — typically develop, not main. PRs target staging/develop first, so commits from earlier today or this week may not be on main yet. Confirm the branch with git branch --show-current if unsure, or git log <branch> --no-merges to see what's actually there.
- For closed weeks (rollup, backfill),
main is usually fine — the week's PRs have merged by then. But if you're rolling up early Monday and the Sunday PRs haven't been promoted yet, fall back to develop.
git log <branch> --no-merges --pretty=format:"%H|%ad|%s" --date=short --since=<week-start> --until=<week-end> — --no-merges filters out PR merge commits, which carry no useful subject. Range is usually <last-rollup>..HEAD or explicit --since/--until.- Sanity-check: if the user just mentioned a specific commit you don't see in the output, you're on the wrong branch. Re-query.
- Apply curation rules (below). Drop noise, rewrite dev-speak into user language, dedupe revisits.
- Verify net state. Before locking bullets, run the added-then-removed check (Verify the net state, not the commit stream): drop any Added/Improved bullet whose feature was torn out again before HEAD. This is a distinct pass from step 4 — the
feat commit passes curation on its own; only cross-checking it against later teardown commits reveals it should go.
- Group into Added / Fixed / Improved.
- Show the result. Wait for approval before touching files.
Append mode
After the user approves drafted bullets, write them into current in changelog.json and reflect in CHANGELOG.md.
- Run preconditions first. If
current.date is older than this week's Monday → rollup before appending.
- Merge with existing
current.groups — dedupe by title. If a new bullet refines or walks back something already there, edit the existing bullet rather than adding a duplicate.
- Update
current.date to today.
- Update
CHANGELOG.md:
- Bump
## [Current] — week of <date> to <this week's Monday> if the heading is stale.
- Insert new bullets at the top of each group (Added / Fixed / Improved).
- Update
src/data/changelog.json to mirror.
- Do not create tags. Do not update
package.json.
If current already has ~15+ bullets in any one group when you arrive, mention this to the user — it's a signal that a rollup is overdue, not a license to keep appending.
scripts/changelog-append.mjs and npm run changelog:append exist as a convenience for scripted runs, but the manual path (skill curates, then writes) is preferred — it produces better user-facing copy.
Rollup mode
Triggered manually on Monday to close the previous week.
- Determine last week's Monday → version
YYYY.MM.DD.
- Move
current into releases with that version and the Monday date.
- Empty
current (or seed with any commits landed since Monday).
- Update
CHANGELOG.md with the new entry at the top.
- Print the commit/push commands for the operator to run manually.
npm run changelog:rollup exists as a scripted shortcut.
Backfill mode
Given a git range, produce historical weekly entries. Used when seeding from scratch or filling a gap.
- Walk PR merges chronologically:
git log --merges --first-parent main --pretty=format:"%H|%ad|%s" --date=short <range>.
- Group by week (Monday-start). Union all non-merge commits per week.
- Apply curation, dedupe revisits.
- Emit one entry per week with at least one user-visible change. Skip empty weeks.
Pre-PR era (if any): collapse the foundation into one initial release entry, dated the Monday before first-PR week.
The user-facing test
Before keeping any bullet, ask: would a user notice this if they used the app today vs. yesterday, without diffing screenshots? If no, drop it. This is the single most important filter — apply it to every candidate bullet, not just borderline ones.
Concretely, a bullet survives only if it changes something the user can see, do, or feel:
- A new control, panel, shortcut, format, or capability
- A bug they hit (or could hit) is gone
- A workflow that visibly stalled, jittered, or felt sluggish is now smooth in a way they'd remark on
A bullet fails if it only describes:
- Internal mechanics (allocation, caching, dirty-marking, dispatch, refs, props)
- Sub-pixel or single-pixel visual tweaks (centering, alignment, hover color shifts, tiny margins)
- Hit-target widening, drop-zone expansion, ghost-position adjustments — fold into the parent feature instead
- Same-week regression fixes for code shipped in the same week
- Anything the operator only knows happened because they read the diff
When in doubt, drop it. The changelog is read by users browsing "what's new" — they will not appreciate a bullet they cannot perceive.
Frame the story: launch vs. increment
Commit subjects are written from the inside — the author already knows the system exists, so each commit reads like an increment ("add Turkish locale coverage", "localize timeline labels", "fix tts locale strings"). From outside, a series of "increments" can actually be the launch of an entirely new system that didn't exist last week.
Before writing bullets for a feature area, sanity-check the framing:
- Pick the directory or module the commits touch most (e.g.
src/i18n/, src/features/subtitles/, lib/gpu-transitions/).
- Run
git log --diff-filter=A --since=<week-start> -- <path> to see when files in that area first appeared.
- If most of the files were born this week, the area is a launch — write one top-line bullet describing the new capability, and treat the individual "add X" commits as facets of it, not separate bullets.
- If the area has older history, the commits are real increments — describe each that passes the user-facing test.
The user once flagged this directly: an i18n week was drafted as "Turkish language support" + two localization fixes. The actual story was "UI is now translated, with 9 languages." The Turkish commit looked incremental because the operator was thinking from inside the feature; the user reading "What's New" had no prior context that translations existed at all.
When you spot a launch:
- Lead with the capability ("Translated UI in 9 languages — …", "Subtitle editing on the timeline", "Pen tool for masks").
- List the supporting facets only if they're not obviously implied (a language picker is implied by "translated UI"; ASS subtitle support inside subtitle editing might be worth its own bullet).
- Resist the urge to itemize every commit just because they're all in the Added group.
Verify the net state, not the commit stream
Commits are a stream of deltas; the changelog describes the net difference between the last release and HEAD. A feature added on Tuesday and ripped back out on Thursday is, to the user, as if it never shipped — there is nothing to announce. But the naive draft still emits an "Added" bullet, because the feat(...) commit is sitting right there in the log and nothing downstream cancels it in the text. The removal happened in the code, not in the commit subject you're reading.
Before finalizing any Added or Improved bullet, confirm the thing it describes still exists at HEAD:
- Find the teardown commits in the window. Scan subjects for
/\b(remove|delete|drop|revert|roll ?back|back out|undo|kill|rip out|disable)\b/i, and list file deletions with git log <branch> --no-merges --diff-filter=D --since=<week-start> --until=<week-end> --name-only.
- Pair each candidate Added bullet against them by feature area, not literal path. A feature can be dismantled by deleting a route, a menu entry, a flag, or a registry line without deleting the file that introduced it. Match on what the user would touch, not the filename.
- Added and removed within the same window → drop the bullet entirely. Not Added, not Fixed, not Removed. It nets to zero for the user; announcing it and then never shipping it is worse than silence.
- When unsure whether it survived, verify against HEAD directly — don't trust the
feat commit alone. git cat-file -e HEAD:<path> for a file's existence, or grep HEAD for the symbol / route / UI label the feature exposes. If it's gone from HEAD, it's gone from the changelog.
This generalizes the "reverts paired with a re-fix" rule below. It also catches the cases that rule misses: an add later removed with no replacement, an add superseded by a differently-named replacement (log only the survivor, worded as the survivor), and an experiment merged in one PR then reverted in another.
The one exception — removing something that already shipped. If the removed feature was in a prior release, the user has been using it, so its removal is itself a user-facing change: they had it yesterday, it's gone today. That is worth announcing (and any migration note that comes with it). The JSON schema has no Removed group, so surface it to the user and let them decide wording and placement rather than silently dropping it or inventing a group. The added-then-removed-same-window case above is the opposite: the user never had it, so nothing is announced.
Curation rules
Drop
- Merge commits (
Merge pull request, Merge branch)
chore(...) — including chore(release)ci(...)test(...) unless it documents notable test infra changesrefactor(...) when the scope is internal (stores, types, utils, deps adapters, chunk splits)deps / deps-dev bumps unless major-version bumps with user-visible impact- Follow-up fixes — commits whose message matches
/address.*(review|PR|findings|feedback)|code review|follow-?up|fix.*(lint|typecheck|build|pre-existing)/i AND land in the same week as a parent feature. Roll into the parent bullet silently.
- Same-week regression fixes — fixes for bugs introduced by other commits in the same week. The user never saw the regression; only the final state matters. Fold into the parent feature bullet, do not add a separate "Fixed" entry.
- Visual polish micro-tweaks — recentering, hover-color changes, padding/margin nudges, alignment fixes that no user would notice as a discrete improvement. If a feature has 3+ such tweaks, the parent feature bullet absorbs them.
- Hit-zone / drop-zone / ghost-position adjustments — fold into the parent drag/drop feature. Never their own bullet.
- Internal perf on this week's new work — perf commits that optimize code shipped earlier in the same week. The user experiences the feature once, smoothly. No separate "Improved" bullet.
- Reverts paired with a subsequent re-fix in the same week — skip both, keep only the final correct implementation.
- Added-then-removed within the window — any feature introduced and then torn out before HEAD (with or without a replacement). See Verify the net state: it nets to zero, so drop the Added bullet. The exception is a feature that shipped in a prior release and is removed this week — that removal is user-facing; surface it.
- "Update src/..." auto-subject merges (GitHub web-edit artifacts)
- Revisits: if the same feature is improved multiple times in one week, dedupe to one bullet describing the final state.
- Duplicates worded differently — drag overlays sticking, drag overlays hijacking lanes, stale drop overlays are all "dragging works better now." One bullet, not three.
Keep and rewrite
feat(...) — when the feature is user-perceivable (a new control, mode, format, shortcut, panel). Skip internal feat(...) that only changes plumbing.fix(...) — only if the bug was user-observable (rendering glitch, playback hitch, data loss, crash, wrong output) AND shipped to users before the fix. Skip fixes for code never released, or shipped and corrected in the same week.perf(...) — only if the user would describe the change in their own words ("playback is smoother", "scrolling doesn't lag anymore"). Skip perf the user can't perceive without instruments.
Rewrite style
Commit messages are dev-speak. The changelog is user-facing. Rewrite every bullet.
| Commit subject | Changelog bullet |
|---|
feat(timeline): add Alt+C as alternate split-at-playhead shortcut | Split clips at playhead with Alt+C |
perf(filmstrip): fill zoom gaps with cover frame background and full-set fallback | Smoother filmstrip rendering when zooming the timeline |
fix(preview): retry video with fresh blob URL on stale-blob load errors | Preview no longer fails when media blobs expire |
feat(storage): migrate to workspace folder via File System Access API | Projects now live on disk in a folder you choose, not hidden browser storage |
Rules of thumb:
- Lead with the verb of the user experience, not the code change.
- Drop internal names (stores, modules, workers) unless the user knows them.
- If a bullet is only meaningful to developers, drop it.
- Aim for ≤12 words per bullet.
- For weekly entries, prefer thematic phrasing over commit-level phrasing ("Trim tools with smart zone detection" rather than listing each tool variant).
- Sentence case: write every bullet starting with a capital letter. The dialog has a render-time safety net that capitalizes the first character, but
CHANGELOG.md ships the raw string to GitHub — fix the source, don't rely on the UI fallback.
Grouping
Within each weekly entry, group into:
- Added — new user-facing features
- Fixed — bugs the user actually hit (in shipped builds)
- Improved — performance or polish wins the user would describe in their own words
Skip any group with zero entries. It is normal and good for "Improved" to be empty. If you cannot phrase an Improved bullet in a way the user would say it out loud ("playback is smoother", "exporting is faster"), it does not belong here. Do not pad the section with internal perf to make the week look productive.
File formats
src/data/changelog.json
Matches types in src/data/changelog-types.ts:
export type ChangelogGroup = 'added' | 'fixed' | 'improved';
export type ChangelogItem = {
title: string;
};
export type ChangelogEntry = {
version: string;
date: string;
subtitle?: string;
groups: Partial<Record<ChangelogGroup, ChangelogItem[]>>;
};
export type ChangelogFile = {
current: ChangelogEntry | null;
releases: ChangelogEntry[];
};
The What's New dialog renders one entry per group with title-only bullets — no scope tags, no highlights section. Make every title carry its own context (a user reading it cold should know what changed).
CHANGELOG.md
# Changelog
All notable changes to FreeCut. Versioning follows weekly CalVer: `YYYY.MM.DD` = the Monday of the release week.
## [Current] — week of 2026-04-13
### Added
- ...
### Fixed
- ...
## [2026.04.06] — week of 2026-04-06 to 2026-04-12
...
No **Highlights** block. No scope prefixes on bullets.
Do not include PR links in weekly entries. Weekly entries aggregate many PRs; PR links add noise.
When in doubt
- Fewer bullets > more bullets. A wall of text is worse than missing a tiny fix.
- If a week is 100 commits but they all revisit the same 3 features, produce 3 bullets.
- If a commit scope is new and you don't know if it's user-visible, check what files it touches. UI components and public APIs = user-visible; stores/utils/tests = usually not.
- A
current entry past ~15 bullets means it's time for a rollup.
- When the user asks to "update the changelog" with no other context, default to draft mode — show curated bullets first, write nothing until they approve.
