Reference analyzer
Location: /studio/[project]/reference
An analysis tool that mines directing devices from video references. It extracts frames (8-32) from an uploaded video file or a YouTube link, and a vision model analyzes each frame for shot size, angle, composition, color grade, lighting, camera motion, transition and on-screen text, producing per-frame prompts and an overall edit-style profile. Results auto-save to IndexedDB for reuse from a project/global library, and 'Apply as pack' registers the style as an inline picker available in every prompt box.
- 1Pick a frame count (8/12/16/24/32), then upload a video file or paste a YouTube link and click Analyze.
- 2Watch the extract-then-analyze progress (it keeps running across tabs and can be cancelled).
- 3Read and copy the overall summary and style phrase in the edit-style profile at the top.
- 4Inspect per-frame cards for shot/angle/composition/color chips and the English prompt, copying what you need.
- 5Click Apply as pack to register the analysis as a prompt-picker tool (ref-style).
- 6Open or delete saved analyses from the library, scoped to the project or all projects.
Vision analysis: prefers a local CLI agent (Codex / Claude Code, no key), falling back to a BYOK vision provider (OpenAI / Anthropic / Google) or a custom LLM endpoint. Frame extraction uses /api/owntent/video-frames and client-side extraction.
No BYOK key is needed if a local Codex or Claude Code CLI agent is logged in; it handles the vision analysis. Results auto-save, so apply strong references as a pack to reuse them across other prompt boxes.