A four-skill suite for producing explainer videos with agents
Retroactive log - reconstructed from commit history and session notes.
We shipped a four-skill suite that codifies the explainer-video pipeline so it can run repeatably with agents instead of being reconstructed from scratch each time.
The four skills cover the full arc from script to publish. /video-script takes a topic or an existing article and produces a six-beat house-format script with visual notes and named anchors. /video-build takes an approved script and voice-over, scaffolds a videos/<slug>/ directory in the video repo, runs Whisper forced-alignment to produce a timing.ts file, and renders. /video-package takes the final MP4 and produces SRT captions, YouTube chapters, a youtube.md metadata file, and thumbnail candidates. /video-publish uses Playwright to upload to the channel and opens a cross-promo PR in the content repo.
Each skill ships as a full triplet: the canonical SKILL.md in .agents/skills/, a dispatcher in .claude/commands/, and a Gemini mirror. All four pass skill-review --all with zero errors (four warnings and five informational findings, all pre-existing from other skills in the repo).
The companion house-style doc captures the brand tokens, the 30fps + legibility floor rules, the six-beat structure, the Whisper forced-alignment pattern, and the publish checklist. It also documents the ElevenLabs voice-over key gap - the pipeline has a slot for a VO key but the key itself is not yet provisioned, so the current flow requires a manual VO hand-off before the build skill can run.
This built on top of a companion PR in the video repo (#1) that restructured the project layout to shared src/ plus per-video videos/<slug>/, with slug-parameterized scripts so each video is isolated.
What surprised us: the Whisper forced-alignment pattern - producing a timing.ts from a voice-over file so captions sync to speech - required more consideration than the upload-to-YouTube part. Most of the complexity in the package skill comes from getting the timing right, not the platform API.
What’s next: provision the ElevenLabs key and run the full pipeline end-to-end on the first video.