Intro Editor

Intro Editor

Turn a bare talking-head recording into an intro edited like Samin's reference video: the camera audio is the spine, and a dense visual layer sits on top of it. The viewer should almost never look at a static talking head for more than ~15 seconds.

Mental model

Two layers, always:

1. Audio spine — the original camera take's voice, untouched and continuous. Never cut or reorder audio within a beat; the edit is visual.
2. Visual layer — what's on screen switches constantly between (a) the camera itself at alternating zoom levels and (b) full-screen inserts with a small PiP bubble of the speaker for continuity.

Every sentence the speaker says gets classified: does it name a tool or result (→ screen recording), state a number or claim (→ stat graphic), tell a story (→ illustrated scene), lay out structure (→ agenda card), or carry pure energy (→ stay on camera, maybe punch in)? The full grammar with timestamped examples from the reference video is in references/style-spec.md — read it before planning any edit.

Inputs to collect

• The raw video: local file path, YouTube/Drive link, or an existing Jockey asset. If the user has separate b-roll / screen recordings, get those too.
• Which part is intro material (if unclear, propose a scope after analysis — usually the first continuous "setup/promise" section of the talk).
• Where finishing happens: ffmpeg draft render (default) or an edit plan for their NLE.

Pipeline

Work through these steps in order. Steps 1–2 can overlap (frame analysis is local and doesn't need Jockey).

1. Ingest

Get the video BOTH locally (for ffmpeg frame work and assembly) and into a Jockey knowledge store (for transcript-level understanding and clip search). Follow references/jockey-pipeline.md — it has the exact validated commands, including the Zo relay trick for local files (Jockey cannot fetch YouTube URLs or local paths) and polling until the item is ready.

2. Analyze the raw footage

Two passes, different tools:

• Visual pass (local, immediate): extract timestamped contact sheets with ffmpeg (command in references/assembly.md) to see framing, existing cuts, and where the speaker is just talking to camera.
• Speech pass (Jockey): pull a timestamped beat map — every claim, number, story, tool mention, and structural statement, with start/end times. Use jockey_query with a json_schema so beats come back structured. Query patterns are in references/jockey-pipeline.md.

3. Choose the intro scope and cold open

The reference style front-loads proof: the first ~15 seconds are a rapid montage (2s cuts) of end results — charts, numbers, the product working — pulled from later in the video or from the asset library, playing under the opening lines. Identify: the hook sentence, the promise/roadmap section, and 3–6 "proof moments" from anywhere in the footage that can be harvested for the cold open.

4. Map beats to visual treatments

Go beat-by-beat through the intro transcript and assign a treatment from the grammar in references/style-spec.md. Enforce the pacing rules there (max camera hold ~15 s, insert lengths 2–8 s, return to camera between insert clusters). The output of this step is the Edit Decision List (EDL) — use the exact format below.

5. Gather and generate assets

For each planned insert, resolve a source in this priority order:

1. Harvest from the same video — screen recordings and demos later in the footage (find via jockey_search, verify with a frame extract).
2. User's existing library — other Jockey stores or local folders they point at.
3. Generate — illustrations (pixel-art / comic / vintage styles per the spec), stat graphics, agenda cards, and title cards. Use available image/video generation tools; the spec has style prompts. Anything not generatable becomes an asset request list for the user (e.g. "record 10 s of the dashboard showing X").

Never leave a beat with a dangling treatment: every EDL row ends resolved (file path) or requested (in the asset list).

6. Deliver the EDL

Save intro-edl.md (human-readable) plus intro-edl.json (for assembly) next to the project. Every row:

| # | Base in–out | What's said (short) | Treatment | Visual description | Source | Status |

Treatment is one of: camera camera-punch screen-rec stat-graphic illustration agenda-card title-card overlay-only. JSON schema in references/assembly.md.

7. Assemble a draft

Render with ffmpeg following references/assembly.md: base track cut to the intro, inserts overlaid full-screen with the PiP bubble composited bottom-right, punch-ins via crop+scale, title cards as generated stills. Output intro-draft.mp4.

8. QC before showing the user

Make a contact sheet of the rendered draft and read it. Check: no camera hold over ~15 s, PiP present on every full-screen insert, cold open starts on proof not on the speaker, audio spine unbroken. Fix before delivering. Show the user the draft plus the EDL and the asset request list.

Quality bar

The edit fails if any of these are true: the first shot is the bare talking head; any insert plays without the PiP bubble; a stretch of bare camera exceeds ~20 s; an insert contradicts what's being said (visual must match the sentence under it); the audio has gaps or doubled words at cut points.