video-shotcraft vs Remotion's official skills: same brief, both rendered
Two Claude Code skills promise to turn an agent into a video studio — a 164MB community production kit and the framework vendor's own 10-skill pack. Instead of comparing READMEs, we gave both the identical five-second brief, followed each skill's own methodology to the letter, rendered both results, and then wrote up what we would fix in each repo. This is the comparison a directory cannot write.
Tested 2026-07-25
The two contenders
| Metric | video-shotcraft | Remotion official skills |
|---|---|---|
| Repo | Vincentwei1021/video-shotcraft | remotion-dev/skills |
| Stars | 400 as of 2026-07-25 | 4,073 as of 2026-07-25 |
| Commit tested | a87bf81 | 0dd76fa |
| Install | npx skills add Vincentwei1021/video-shotcraft -a claude-code -y | npx skills add remotion-dev/skills -a claude-code -s '*' -y |
| Install time | 11.9s measured 2026-07-23, same CLI path | 2.5s measured 2026-07-23, same CLI path |
| Payload on disk | 164MB | 0.59MB |
| Files installed | 686 | 103 |
| SKILL.md | 1 file(s), 144 lines, 407 words — Chinese (frontmatter description included) | 10 file(s), 591 lines, 2446 words — English |
| Audio assets bundled | yes | no (guidance only) |
The size gap is the first real finding: one side is a kit (shot recipe cards, a finished template, bundled SFX with attribution), the other is rules (ten instruction files totalling under a megabyte, no assets). Neither is wrong — they are different theories of what an agent needs. The rest of this page tests those theories.
The brief (identical for both)
Small on purpose: typing animation, a spring, one composition — enough to force each skill to reveal how it actually directs an agent, without a day of rendering.
How each side told the agent to work
video-shotcraft: pick a card, inherit its taste
The skill's workflow is opinionated before a single frame exists. Its SKILL.md hard-gates the
session: introduce the bundled Ink Press template first and ask whether to use it or go
free-form. For our free-form brief, its methodology routes you into
references/shots/ — we used the typewriter card, and it handed over
implementation-grade specifics: characters revealed every 2 frames, a block cursor on an
f%12<6 square wave that stays solid while typing and dies two blinks after, an
overshoot bezier lifted from the template's own tuned component for the check pop, and an
explicit "hold ≥50 frames of stillness" ending. The agent inherits a cinematographer's
defaults.
Remotion official skills: correct by construction
The official pack routes through remotion-create into
remotion-markup's rulebook. Its timing file states a clear default — prefer
interpolate() with a bezier over spring() unless the user
explicitly asks for physics; our brief said "with a spring", so a spring it was. Rules
covered clamping extrapolation, keeping animation inline in the style prop so Remotion Studio
can write edits back to code, and individual transform properties. What the rulebook never
offers is an opinion about how the result should look.
What actually rendered
Frame 75 from each side, unretouched:

video-shotcraft — shot-card path

Remotion official skills — rules path
Same brief, same fonts available, same renderer — and the difference is exactly the two theories made visible. The card path produced a composition: centered title, a badge (green disc, white check) scaling in below it, because the card and the template component it borrows encode design decisions. The rules path produced a correct implementation: the text types on, the spring springs — as a bare glyph sitting wherever the agent happened to put it. Every API rule was followed; no one was directing.
The numbers
| Metric | video-shotcraft | Remotion official skills |
|---|---|---|
| Project setup | 7.6s npm install inside bundled template/, measured 2026-07-23 | 29.8s manual minimal project npm install (fallback after create-video scaffold failed, see failures) |
| Still, frame 75 | 5.8s | 6.6s |
| Full 150-frame render | 13.8s | 13.5s |
| Output MP4 size | 250,402 bytes | 251,881 bytes |
| Failures on the way | 1 | 2 |
At render time the two are indistinguishable — same engine, same walls, near-identical timings and file sizes. Everything that separates these skills happens before the render: what the agent is told to care about, and what it ships to your disk to say it.
What broke, verbatim
On the video-shotcraft path
On the official-skills path
And one wall both sides hit identically, because it belongs to Remotion itself, not to either
skill: on a headless Linux box with no system Chrome, the renderer tries to download a headless
shell from remotion.media — blocked in sandboxes and CI allowlists. The fix that
works for both: point --browser-executable at any chrome-headless-shell binary
(Playwright's cache has one), and add --concurrency=1 on low-core machines. Full
failure chain on the video-shotcraft test page.
A note on the star counts
| Skill | Star readings (GitHub UI/API, dated) |
|---|---|
| video-shotcraft | 421 (2026-07-22) → 564 (2026-07-23) → 400 (2026-07-25) |
| Remotion official skills | 3,677 (2026-07-23) → 4,073 (2026-07-25) |
We record stars with dates because they move — and not always up. video-shotcraft's count fell between our test and this comparison while the official pack's rose. We log the readings and draw no conclusion beyond the obvious one: star counts are a weak signal, which is why this site measures installs instead.
What we would fix upstream
A deep comparison should pay its subjects back. These are concrete, evidence-backed suggestions we are filing with both authors. The first is up: remotion-dev/skills#24 (the hub-links finding below, with full repro). (One correction, logged in the changelog: we first described the hub's dead links as files missing from the pack; checking the repo source showed the truth is subtler — an install-time rename/rewrite that misses exactly those six links.)
To video-shotcraft
1. Provide an English SKILL.md variant
SKILL.md (144 lines) and its frontmatter description are Chinese-only while the README is English-first; the frontmatter description is what the agent matches against English task phrasing, so activation and instruction-following pay a translation tax for the repo's primary audience.
2. Slim the default install payload
164MB / 686 files land in .claude/skills/ per project; gallery/ and demos/ are human preview material the agent never reads at production time. An optional download or a documented -g recommendation would cut most of it.
3. Document the constrained-environment render path in SKILL.md
On machines without Chrome, Remotion downloads a headless shell from remotion.media — blocked in sandboxes and CI allowlists (403 on 2026-07-23, DNS-blocked on 2026-07-25). One paragraph on --browser-executable plus --concurrency for low-core machines would save users the hours it cost us.
To remotion-dev/skills
1. Six hub links break after a skills-CLI install (rename/rewrite misses them)
In the repo, sub-skills live as <skill>/REFERENCE.md and the hub links to them. On npx-skills install, each REFERENCE.md is renamed to SKILL.md and cross-links of the form ../<skill>/REFERENCE.md are rewritten to match — but the hub's six prefix-less links (<skill>/REFERENCE.md) escape the rewrite. Post-install, the hub's entire routing is dead: grep shows the six REFERENCE.md links in the installed hub, and no REFERENCE.md file exists in the installed layout. (Corrected 07-25: an earlier version of this page said the files 'do not exist in the pack' — they exist in the repo; they are dead only after install.)
2. Add a consumer-facing README
The README says in full that this is an internal package with no documentation, yet the repo has 4,073 stars (2026-07-25) and a first-class skills-CLI install path. Ten lines listing the skills and routing to remotion.dev/docs would fix first contact.
3. Make the create-video scaffold proxy-aware
npx create-video@latest resolves registry.npmjs.org and github.com directly and dies with EAI_AGAIN behind a DNS-restricted proxy, while plain npm install succeeds in the same shell — the scaffold ignores the npm proxy configuration that is demonstrably present and working.
Verdict: which one, when
Choose video-shotcraft when the deliverable is a product promo and you want it to look directed: the cards carry taste (timing, easing, composition, even sound design with bundled, attributed SFX), and the Ink Press template is the fastest path from screenshots to a finished-looking film. Budget the 164MB and read the fixes above if you run headless.
Choose the official pack when you are building your own visual language or anything that is not a promo — captions, audio visualization, SaaS embedding, interactive Studio work. It is 300× lighter, English-first, API-correct by construction, and covers far more of Remotion's surface. Just bring your own art direction, because it will not offer any.
Run both if you can: they do not conflict, and the combination — official rules for correctness, shot cards for taste — is what actually resembles a studio.