Raw Takes to Published Video: the approved cut

Vibe editing: the cut

Turn raw takes into a cut that keeps the meaning, then hand it to the engine for graphics. The agent proposes and writes; the renderer computes frames; the author decides and accepts. Approve the story before the design: every change of meaning after the graphics forces a rebuild.

• Routes and tools by material: references/routes.md
• The EDL and the two clocks: references/edl-format.md
• Transcripts and recognisers: references/transcripts.md
• Cutting speech: references/cutting.md
• Captions: references/captions.md
• Sound: references/audio.md
• Inserts, layouts, 9:16, other languages: references/inserts-and-layouts.md
• Screen recordings: references/screen-recordings.md
• Folder, brief and project.md: references/project-memory.md
• Notes the author gives: references/edit-requests.md
• Data, licences and money: references/data-and-licences.md
• Plan format: references/plan-format.md
• What this repository measured: references/measured-behavior.md
• Versions and dates: references/versions.md
• Citations: references/sources.md

Contract

• Phase 1 is read-only. The toolchain report, the media inventory, reading transcripts that exist, and writing the plan change nothing and send nothing anywhere.
• No cut, render, upload, install or paid call without approval of item IDs (references/plan-format.md). "Make it shorter" before a plan exists is a request for the plan.
• Originals are never changed. Working copies go to work/, renders to renders/, deliveries to exports/. edl.py render refuses to write over a source or an existing file without --force.
• Two clocks. Source time for takes, output time for everything a viewer sees; convert with edl.py map and edl.py retime, never by hand.
• Cite or label. A claim about a tool cites references/sources.md or a Measured section; versions and prices come from references/versions.md with its date; a judgement of taste or meaning is labelled as such.
• Report verification honestly. A frame check misses rhythm; a script misses meaning. Say what was watched, listened to, scanned — and what was not.

1. Preflight

python3 scripts/toolchain_report.py --format markdown
python3 scripts/media_probe.py originals/ --format markdown

• The toolchain report says what this machine can do: frame-accurate renders, burning captions (many FFmpeg builds cannot, Measured M7), loudness measurement, local speech recognition, HyperFrames, Remotion; and which API key names are set (never their values).
• The inventory flags what changes the cut: variable frame rate (MED001), rotation (MED002), HDR (MED010), mixed frame rates or audio formats (MED004, MED005); rules in CHECKS.md in this skill's repository.
• Read the project's own instructions and project.md if they exist.
• Write the brief, or ask for what it lacks: viewer and takeaway, delivery, must-stay, must-not, cloud (references/project-memory.md).

2. Route

Pick the route from the material and the deliverable (references/routes.md). This skill owns the cut: takes, joins, the EDL, the render, captions and sound on the new timeline. Graphics belong to the engine's skills: HyperFrames (hyperframes, general-video, hyperframes-core) or Remotion (remotion-best-practices). Captions or overlays on footage that must not change belong to embedded-captions or talking-head-recut.

3. Transcripts

• Word timestamps per source, made once, kept in transcripts/ (references/transcripts.md).
• Recognition in the cloud sends audio out and costs money; local recognition needs an install. Both wait for approval, with the provider, the cost and the data named in the plan.
• Correct names and numbers against a glossary before planning any cut.

4. Plan the story

1. Table of the material: source, in–out, verbatim phrase, sound, what is visible, possible role.
2. Three openings, each with what the viewer understands in the first seconds.
3. Candidates: python3 scripts/edl.py candidates --words transcripts/<take>.json — pauses, repeats and fillers to consider, never to cut automatically (references/cutting.md).
4. The story in beats, then edit/edl.json: every clip with a reason (references/edl-format.md).
5. python3 scripts/edl.py check edit/edl.json --probe, table and cuts --probe: the clips with both clocks, and every removed interval for the author to read.
6. Doubtful cuts go into edit/edl_b.json, not into version A.

5. Render the draft

python3 scripts/edl.py render edit/edl.json -o renders/draft_v1.mp4 --size 640x360 --preset veryfast

• Every clip is re-encoded on frame bounds, with 10 ms audio fades, scaled to one size (Measured M1, M3, M4). A stream-copy cut is never frame-exact (Measured M2).
• The first draft has no music, captions or effects: speech and picture only.
• Listen to every join on headphones; look at the frames either side. Fix with narrow notes (references/edit-requests.md) and a new EDL version.

6. Captions, sound, inserts

• Captions: edl.py retime moves word times onto the output timeline; captions.py build groups them; captions.py check applies the limits (references/captions.md).
• Sound: speech first, then music with ducking, then effects; measure loudness against the brief's target and set loudnorm targets explicitly (references/audio.md).
• Inserts: mark them on the output timeline (edl.py map), approve a still frame, then build them in the engine; edl.py hyperframes and edl.py remotion hand over the cut (references/inserts-and-layouts.md).
• Other formats: a 9:16 version is a new layout, not a centre crop.

7. Hand over

• Run the review before calling the video done: the vibe-editing-review skill, or at least python3 scripts/qa_scan.py with the brief, the EDL and the captions.
• Update project.md: approved EDL version, last good render, commands, assets and licences, open defects.

Workflow

Phase 1 — Recommend

1. Preflight: toolchain report, inventory, brief, project instructions.
2. Plan with references/plan-format.md: sources and conversions; the story and EDL v1; removed intervals; doubtful cuts; captions, sound and inserts; data, money and licences; what cannot be verified here.
3. Ask which item IDs to apply.

Phase 2 — Apply and verify

1. Apply approved items only, one level at a time: story, pace, text, picture, sound.
2. After each change, render or preview the same range with a second of context, and listen.
3. Re-run edl.py check and captions.py check; run qa_scan.py on every draft that goes to the author.
4. Update item statuses and project.md; report what was not verified.

Guardrails

• Do not cut, render, upload, transcribe in the cloud or install anything in Phase 1.
• Do not change or re-encode files in originals/.
• Do not stream-copy a cut that must be frame-exact.
• Do not place captions or inserts by source time.
• Do not remove every pause with one threshold, or synthesise a missing word without the author's decision.
• Do not print a key or a token in the chat, a log or a file.
• Do not add music, fonts or footage without a recorded source and licence.
• Do not take over the engine's ground: composition code belongs to the HyperFrames or Remotion skills.