video-shotcraft

An agent skill that turns Claude Code or Codex into a motion-design studio: point it at a product and it storyboards, animates and sound-designs a cinematic promo with Remotion. The library is the point — 157 shot recipe cards under ten functional categories, each one documenting purpose, energy, suggested duration, parameters, implementation notes and known pitfalls, each paired with a TSX implementation carrying the real easing and timing values and with a motion preview you can filter in the online gallery. On top sit a validated 36.2-second template film (1920×1080, 30fps, ten shots, paper-ink-amber), 149 sound effects filed under sixteen scene categories, five music beds, a 2.5D page camera for real page captures, an exporter that turns the finished film into an editable JianYing draft, and a browser workbench that reopens the delivered film as shot, transition, caption and sound tracks. Install is a symlink into the agent’s skills directory or one npx command; Apache-2.0, about 183 MB.

Who built itThe repository is published under the GitHub account Vincentwei1021, and the README signs off as Vincent. Its 101 commits are signed under two names — Wei Yihao, from an amazon.com address, on 47 of them, and Yihao, from an address linked to the account, on 36 — which together account for 83 of the 101. Seventy-one of the commits carry a co-author trailer, and 62 of those name a Claude model: Fable 5 on 47, Opus 5 with a million-token context on 12, Fable 5.1 on 2 and Claude Code on 1. Nine contributors appear in total; after the author, the largest has six commits.

1. 01

   Eleven weeks, three monthly batches, and a card count that never stopped moving

   The repository was created on 2026-07-19, and its first commit puts the whole of version one into a single message: the card library, the demo implementations, the finished-film template, the assets and the gallery. 101 commits follow, and their distribution is the shape of the project — 66 in July, 25 in August and 10 in September, the last of them dated 2026-09-28 and adding Russian to the workbench interface. The published numbers move in batches rather than continuously. The August note announces 48 new cards and says the library grew from 104 to 152 cards and 209 previews, distilled from 209 candidate motions through eight rounds of frame-by-frame review against reference footage. A contribution on 2026-09-01 records 152 cards and 209 previews becoming 155 and 212. Today the README banner reads 157 cards and 214 previews while the repository description still says 152 and 209, and the workbench counts 216 demo motions. Around it sit 10,029 stars, 897 forks, 11 watchers and 5 open issues, nine contributors, and two releases that are both media drops — gallery-media on 2026-07-26 and showcase-media on 2026-09-01 — rather than version numbers.

2. 02

   Two hundred and nine candidates, eight rounds of review, forty-eight survivors

   How a card comes into being is stated in the README rather than left to be inferred: 209 candidate motions were put through eight rounds of frame-by-frame review against reference footage, and 48 of them became the August batch. Every survivor arrives as three artifacts — a recipe card under references/shots/<category>/, a native Remotion component at demos/<category>/<name>/<Component>.tsx that is deterministic and driven by a normalized progress value, and a rendered motion preview — and all of them are de-branded on the way in, with neutral stand-in copy and a single swappable accent colour. Where the motions come from is answered in the acknowledgements: twelve named official product films, among them ClickUp, Perplexity, Slack, Notion, Figma, Framer, Bear, Raycast, Pitch, Miro, Superhuman and Loom, were studied for their timing and choreography, and no footage, artwork or brand assets from any of them are in the repository. Contributors keep the same discipline by hand: one pull request adding three cards states that it contains no third-party images, video, logos, product UI or brand names, only deterministic geometry and stand-in copy, and reports the count moving from 152 cards and 209 previews to 155 and 212. The bulk of the work is scripted rather than hand-edited — gallery/sync-from-cards.py and gallery/build-seo.py regenerate the gallery index, a 65,192-byte translation table and the SEO files from the cards, and template/scripts/build-palette-assets.cjs writes a 105,914-byte palette manifest.

3. 03

   Sound was pinned last, and the ordering was learned the hard way

   The 27 KB sound-design document is the most candid process record in the repository, and its first rule is an order of operations: lock the picture, lay the music bed, then pin the effects beat by beat. It was written from a bruising — audio work on the template film only began after roughly thirty rounds of picture changes, and once a soundtrack swap was folded into the same round as a picture rework, the entire effect table had to be re-pinned from scratch. The music bed went through three versions: a warm ambient piano piece by Kevin MacLeod under a CC-BY licence, then a cheerful folk-pop track, then the Mixkit tech-house bed that survived. The first two were rejected by the user within 34 minutes of each other, the second with the verdict that the music and the effects sounded like a video game, and that verdict became a written rule: pick effects by genre rather than by event, so a product promo draws on whoosh, impact, riser, sparkle and transition, and game sound packs are excluded. The library now holds 149 effects across 16 scene and material categories plus 5 music beds, catalogued in a 2026-07-27 audition pass — 36,465 bytes of durations, peaks and suggested pin frames — after a checksum sweep found four pairs of files that were the same download stored twice under two names.

4. 04

   The workbench was ported in from the sibling project, and reviewed in the open

   On 2026-09-04 one pull request brought the browser workbench over from video-talkcraft, the narration-video project in the same series, and wrote down the contract that lets the skill open an editor after delivery. The film is split into shot, transition, caption and sound tracks exactly as authored; the schedule is imported from the film’s own entry point, with an explicit rule that it must not be written a second time; and the demo motions plug in as a library with zero changes to 216 components. Of 218 TSX files, two were skipped because one needs an mp4 asset and the other is a helper, and per-property editing is withheld because the demos are constant-driven — a limit the author recorded in the pipeline document instead of hiding it. The thread doubles as the review record. Five findings were fixed in one commit, among them a manifest key so that a new version with the same total duration still triggers an import, a refusal to reuse a development server the workbench does not own, and a parity script that now exits with its own code rather than reporting a green result it cannot stand behind. A user-reported “frantic flashing” on the first screen was traced to a dozen autoplaying 1080p previews used as thumbnails; thumbnails now rest on a static frame at 45 percent and play on hover, after which zero DOM mutations and zero console warnings were measured, against 28 encoding errors before. A flicker in one scene was located by counting brightness pulses in a screencast, about twenty down to two, and fixed by branching on the rendering environment. An independent review in a fresh context found no blockers.

5. 05

   Two Russian pull requests, one localisation path

   The workbench interface became English by default on 2026-09-23, with a switcher in the top bar and roughly 150 interface strings per language collected in a single translation file, while saved project data is left untouched: track names and card labels are translated at display time, so opening an old project in another language does not rewrite it. Two Russian contributions then arrived. The first localised the interface, translated the card labels and migrated legacy track names inside saved files; the second, eleven days later, carried a full Russian string table built on top of the new language layer. The maintainer closed the first rather than merge two implementations, wrote that the second had conflicts in eight workbench files, and asked for it to be rebased as an extension of the merged layer — same file, same preference key, English still the default, existing behaviour unchanged. The contributor rebuilt it that way and reported that switching language is display-only, with the saved project JSON staying byte-identical, and with the build, the theme tests and a four-frame render parity check passing. The commit the repository now ends on is that Russian interface, merged on 2026-09-28.

6. 06

   A showcase assembled from issue forms, and what the community argues about

   Finished films arrive through a GitHub issue form — 3,066 bytes, with a required statement that the submitter owns the video and consents to it being shown — and are published by an 8,145-byte Python script and its workflow; the bot announces that a work will appear on the showcase page within minutes, and the maintainer converts submissions to 1080p H.264 for web playback. Curation is hands-on rather than automatic: submitters who paste a link are asked to drag the video into the issue, and at least one entry was sent back with the note that it did not appear to use any motion shots from the library. The 30 issues and pull requests on record sample that traffic — a WeChat mini-program promo that named six cards and anchored its cuts at 128 BPM, a browser extension, a patent-coaching tool, an AI trading platform, and a 4K film delivered through a cloud-drive link. One pull request added an optional YouTube field to the submission form and migrated earlier entries that had put YouTube URLs in the wrong field, with 23 tests passing. A theme sidebar arrived the same way, as a feature request plus a draft; by the time it was merged the contributor had cut the diff from 190 files to 42 at the maintainer’s request, dropping generated images and the script that made them. The sharpest thread is a user reporting that the library works almost perfectly with Claude while Codex, Antigravity, Cursor and other agents produced noticeably worse films; the maintainer answered that Codex with a high setting should also work well and that other users report another model. One issue reports flaky continuous-integration tests in a path that does not exist here.