Configurable. Repeatable. Cheap to re-cut.
Sri Studio is the take-home: a text brief becomes a vertical short-form video. Claude Haiku extracts intent, Claude Sonnet drafts a script plan, the user approves in the browser before any paid call fires. On approve, Flux Schnell renders 9:16 scenes, ElevenLabs produces voice-cloned narration with alignment-driven captions, ffmpeg stitches a 1080x1920 MP4, and LangGraph orchestrates the whole pipeline with Langfuse traces. (See the homepage for live runs.)
The take-home was done. I needed demo videos for a separate hackathon submission, and editing them by hand was a day of work I did not have. So I built a pipeline that does it for me, then made the pipeline callable as a service, then wrapped the service for LLM agents. Three extensions, one story.
Silent-first screencast pipeline. Records browser scenes via Playwright, burns captions in BEFORE generating voice, and only fires ElevenLabs after a human-reviewed silent cut. Re-cuts are cheap; voice tokens are not.
The pipeline above, fronted by an HTTP API. POST a project tarball, watch the silent-with-captions preview, POST again to approve and render the voiced final.
A local-stdio MCP server that exposes the HTTP API as four typed tools. An LLM agent (Claude Code) can call helper_render_silent / helper_render_voice without curl, and load SKILL.md as system context with helper_skill.
“This video was generated by the Helper, end-to-end. No editor, no voice actor.”
Each piece was spec'd before it was coded, then the three specs were turned into a single 25-task TDD plan that drove the implementation. Everything is in the studio repo.