Storyboard draft (one-shot pipeline)

Turning a fresh talking-head video into a designed draft is a fixed pipeline. Each stage is its own tool call; skip any stage that get_state's Pipeline line already marks done (transcript done · narration plan done · visual analysis done).

extract_asr → visual_brief → (you label the frames) → submit_visual → plan_brief → (you generate the DraftPlan) → submit_plan → lay_out → per placeholder: compose_block_brief → generate → apply_block

Stage 0 — orient, and offer a theme

Call get_state first. Before the FIRST full-draft run on a project, look at what the video is about and recommend 1–2 best-fitting frames (theme content packs that define the whole design language) from list_frames in one short sentence; let the user pick or skip. When they pick, attach_frame {frame_id} (tokens apply immediately) then read_frame {frame_id} for the playbook — carry its rules into the plan and every block you generate. Don't start the pipeline in the same turn as the question, and never block small edits on it.

Stage 1 — extract_asr (transcript)

Transcribes the spoken audio into timed sentences — the raw material for planning. Covers the main video AND every inserted other-source segment. Slow: it runs in the user's browser and can take minutes — do not retry a call just because it is slow; wait for the receipt. Cheap to re-run later (cached per file). It does NOT add captions and does NOT cut shots.

Stage 2 — footage analysis yourself: visual_brief → submit_visual

This is BYO-brain: you are the vision model, not Pireel's.