| name | video-studio |
| description | The Edit Bay — in-house conversational video studio (VOX-style explainers, documentaries, shorts, LinkedIn video). Transcript-driven cutting, B-roll ladder, motion graphics as code, captions, music, publish-ready packages — all agent-operated, free-first. Use when Farrice says "edit this video/footage", "make an explainer/documentary/short", "cut this recording", "add B-roll/captions/music", "turn this essay/post into a video", or any video production ask. Front door for three modes: talking-head, VO-only, zero-camera (his recorded VO + assembled visuals — NO voice clone, NO TTS narration, his ruling 2026-08-06). Source: Brad Bonanno agentic edit pipeline (extractions/brad-bonanno-edit-bay/), rebuilt on the house stack: WhisperX local + edit_bay.py ffmpeg assembly + HyperFrames/Remotion graphics + /generate (fal) — NO Higgsfield (retired 2026-08-06). |
| version | 1.0 |
| expert | Brad Bonanno (pipeline) × house craft-map masters (direction) |
| domain | agentic video editing + production |
Video Studio — The Edit Bay
Raw inputs (script + any footage) → finished, captioned, 3-format, publish-ready video —
produced conversationally, reviewed by the agent's own eyes, compounding taste every run.
The doctrine (from the extraction — internalize, don't re-derive):
- Give the agent the missing sense. Ears =
transcribe_local.py word timestamps. Eyes = seam-targeted frame extraction of the render. Never conclude "can't" before naming the missing input.
- Controllable inspection > passive comprehension. Review frames at cuts/graphic-edges/caption onsets — never uniform sampling, never a video-understanding model.
- Solved vs fails-by-default. Rough cuts (transcript problem) = solved, spot-check only. Anything visual (overlays, captions, graphics) = fails by default, QA loop MANDATORY.
- Corrections compound or you pay forever. No edit closes until every Farrice note is promoted into the style file's Correction Log.
- Pre-work beats iteration. An annotated script (10 min) saves hours of back-and-forth. Never start assembly from an unannotated script — run
workflows/script-annotation.md first.
Pipeline (six stages — each one agent turn + one CLI call)
| # | Stage | Tool | Cost |
|---|
| 1 | Transcribe | python3 execution/transcribe_local.py <media> --project <slug> --srt | $0 local |
| 2 | Rough cut | agent writes cutlist.json from transcript (silences, filler, bad takes out) → python3 execution/edit_bay.py cutlist-apply --project <slug> | $0 |
| 3 | B-roll | ladder per workflows/broll-ladder.md: own footage → manifest search → free stock → graphics → /generate (fal, cost-gated) | $0 default |
| 4 | Graphics | HyperFrames (overlays on footage) / Remotion (full comps) — graphics render FIRST, enter cutlist as render: shots. Taste gate: load style/graphics-taste.md § named stack (graphics-taste + frontend-design skill + style merge; grace-liu on fog) before ANY comp — never default-style | $0 |
| 5 | Audio | edit_bay.py audio-mix — music bed, VO sidechain ducking, −14 LUFS | $0 |
| 6 | Export + QA | edit_bay.py transcode ×3 presets · qa-probe · workflows/qa-review-loop.md | $0 |
Modes: talking-head (his footage carries) · vo-only (his recorded VO + visuals — the VOX default and current-gear native mode) · zero-camera (same as vo-only; the name means no camera exists, NOT synthetic voice). His voice is always his own — no clone, no TTS narration.
Workspace (per project)
_active/<slug>/05-assets/video/
project.json cutlist.json transcript.json style-overrides.md
footage/ broll/{stock,generated,screen}/ audio/{vo,music,sfx}/
graphics/{hyperframes,remotion}/ captions/ renders/ exports/
Auto-indexed to /assets-board by the existing sweep. Renders version as v01.mp4, v01-frames/, v01-fixlist.json.
Style files (3-level merge — read ALL before any graphic/caption decision)
skills/video-studio/style/video-style-default.md (global floor)_active/farrice-brand/voice/video-style.md (the compounding one — corrections land here)_active/<slug>/05-assets/video/style-overrides.md (project)
Workflows
| Workflow | Use when |
|---|
| produce-explainer | Full production: annotated script → shipped 3-format package |
| script-annotation | Turn a script/essay into an annotated, beat-mapped shooting doc |
| qa-review-loop | Render exists → picky-editor review → fixlist → revise (max 3 passes) |
| broll-ladder | A beat needs a visual you don't have |
Reviewer contract: REVIEW.md · Schemas: schemas/ · Caption presets: caption-styles.json
Hard rules
- NO Higgsfield (retired 2026-08-06). Generated assets only via
/generate + craft-map grammar + fal_budget_guard. seedance-1080p stays HARD-BLOCKED (720p is fine at B-roll/overlay scale).
- Free-first ladder is ordered and mandatory — paid generation is rung 4, quoted before running.
- LinkedIn exports always burn captions (silent-autoplay-safe).
- Stock downloads carry
license + source_url in the manifest (audit trail).
- Every render manifest-indexed; every shipped video has a full version chain.
- Distribution = publish-ready packages; Farrice posts (sends stay human).
- Policy canon:
directives/video-studio-policy.md. Codex parity: everything here is CLI + files on disk — no harness-specific APIs.
Composes with (extend, never rebuild)
skills/remotion-video-creation/ (34 rules; pacing = rules/high-retention-editing.md) · skills/generate/ + craft-map masters (direction layer) · skills/curious-refuge (storyboard floor) · skills/dave-clark (flat-to-cinematic audit) · skills/brad-bonanno-explainer-architecture (what to SAY; this skill is how to CUT it) · skills/jenny-hoyos-shorts (shorts grammar) · fantastic-studio (thumbnails) · hook skills (titles).
