Companion to #1763 (code-breaking defects) and #1753 (licence). These change what a reader decides, without breaking at runtime — which makes them harder to notice.
Wrong values
-
trimBefore / trimAfter are documented in seconds. They are frames. This one contradicts Ref-audio.md inside the same skill, so the skill disagrees with itself.
-
The Lambda ceiling is given as "~2 hours and 5GB". That figure does not appear anywhere in the Remotion documentation. The real constraint is 2048 MB by default (roughly 32 min at 1080p), with a 10 GB cap.
-
"hundreds of Lambda invocations" is 200, and it is a hard cap — not a soft guideline.
-
typography.subtitle is documented as 36. 36 is heading. The subtitle value is 48.
Deprecated or wrong APIs in the examples
-
Patterns.md teaches <Audio> / <Video> with the deprecated startFrom.
-
The only copy-pasteable example in the AI pipeline uses a native <img> instead of <Img>, which is the one thing a Remotion example must not do — a native <img> is exactly the source of frame-tearing that <Img> exists to prevent.
The skill contradicts its own rule
- 14
npx / yarn / pnpm lines across 10 Ref-*.md files, plus npm install and npx remotion render in Workflows/ContentToAnimation.md, which was the only render instruction in that workflow. The skill states bun in its own doctrine, so a reader following the reference contradicts the reader following the rules.
Why these are worth the same attention as the code ones
A wrong unit (seconds vs frames) does not throw; it renders a video with the wrong cut and the author blames their own maths. A ceiling figure that exists nowhere in the docs propagates into cost estimates. In our case the licence line in #1753 caused an agent to cost a pipeline with the wrong model and reach the wrong build-vs-buy conclusion — the reasoning was fine, the file was wrong. These are the same class.
All checked against the installed package and the live documentation; the 12 URLs in the skill were verified to return 200.
Companion to #1763 (code-breaking defects) and #1753 (licence). These change what a reader decides, without breaking at runtime — which makes them harder to notice.
Wrong values
trimBefore/trimAfterare documented in seconds. They are frames. This one contradictsRef-audio.mdinside the same skill, so the skill disagrees with itself.The Lambda ceiling is given as "~2 hours and 5GB". That figure does not appear anywhere in the Remotion documentation. The real constraint is 2048 MB by default (roughly 32 min at 1080p), with a 10 GB cap.
"hundreds of Lambda invocations" is 200, and it is a hard cap — not a soft guideline.
typography.subtitleis documented as 36. 36 isheading. The subtitle value is 48.Deprecated or wrong APIs in the examples
Patterns.mdteaches<Audio>/<Video>with the deprecatedstartFrom.The only copy-pasteable example in the AI pipeline uses a native
<img>instead of<Img>, which is the one thing a Remotion example must not do — a native<img>is exactly the source of frame-tearing that<Img>exists to prevent.The skill contradicts its own rule
npx/yarn/pnpmlines across 10Ref-*.mdfiles, plusnpm installandnpx remotion renderinWorkflows/ContentToAnimation.md, which was the only render instruction in that workflow. The skill states bun in its own doctrine, so a reader following the reference contradicts the reader following the rules.Why these are worth the same attention as the code ones
A wrong unit (seconds vs frames) does not throw; it renders a video with the wrong cut and the author blames their own maths. A ceiling figure that exists nowhere in the docs propagates into cost estimates. In our case the licence line in #1753 caused an agent to cost a pipeline with the wrong model and reach the wrong build-vs-buy conclusion — the reasoning was fine, the file was wrong. These are the same class.
All checked against the installed package and the live documentation; the 12 URLs in the skill were verified to return 200.