Edit: one plan, one command
The Editor turns the user's recording into a video in two parts. First the speaker cut: the best takes, cut at
pauses, graded. Then the edit plan (edit/plan-vNNN.json): every insert, its words and cues, the sound. The
workbench crewcut render builds the plan. Nothing in this skill holds a colour, a font or a level: those come from
the style pack (crewcut brand show). Taste lives in brand/rules/edit.md and brand/rules/cut.md.
Read first
| You want to | Read |
|---|---|
| cut the recording into the speaker cut | references/record-to-cut.md (sound to the end, split files, last full take, double starts, cuts at pauses, re-shoot, grade, show and lock) |
| write the plan, choose or build a scene, show a recording | references/plan-and-scenes.md (word anchors, count the mix, form per statement, cues, new 2D scene, reuse, recordings) |
| make music, set levels, or rent a GPU | references/sound-and-gpu.md (music along the story, chosen by measurement, levels, GPU pitfalls, the cost card) |
| build a 3D scene | skill 3d-scene |
| check before showing | skill cut-check (run by the Critic, never by you) |
Scene names and their data keys: templates/2d/BLOCKS.md and templates/3d/README.md. Do not copy key lists here.
How a version is made
- Speaker cut (
references/record-to-cut.md):crewcut ingest,crewcut transcribe,crewcut cut …. Runpython3 skills/cut-check/scripts/cut_check.py <decisions>on it and fix every ERROR before you build on it. - Words of the final cut:
crewcut transcribe audio/voice-vNNN.wav→work/words-vNNN.json. All plan times come from these words. - Count the mix, then list the spots (section below). Send the list as a card before the big render:
crewcut card propose --episode $CREWCUT_EPISODE --step edit --kind mix --title "Inserts for v002" --file edit/mix-v002.md. Keep building while it is open; the card is a look, not a gate. - Write the plan with a small generator script (
edit/make_plan.py) that sets every time from a word anchor. Never type seconds by hand. Schema:docs/architecture.md, "Edit plan". Values leftnullcome from the pack. crewcut render plan edit/plan-vNNN.json: lists every insert with frames and cues. Read it. An insert of half a second is a wrong anchor.crewcut render 2d edit/plan-vNNN.json: 2D inserts, locally.- 3D:
crewcut render probe edit/plan-vNNN.json <id>gives one still frame at half size. Look at it next to the reference before anything else (skill3d-scene). Then the cost estimate and the card (section "Rented GPU or local"), thencrewcut render 3d edit/plan-vNNN.json --rented --gpus Nor--local. crewcut render cut edit/plan-vNNN.json: assembly with the music bed, hits, typing, gaps, ducking and the loudness target from the pack. The output is unchecked. Tell the Showrunner: "built, unchecked, do not show".- Hand the version to the Critic
(
crewcut job add --episode $CREWCUT_EPISODE --identity critic --step edit "check plan-vNNN"). Only after both checks show 0 ERROR: the preview goes toexport/, the Showrunner shows it as a card with the file (crewcut card propose … --kind preview --file export/<id>-vNNN-preview.mp4), and right after showing:crewcut render shown edit/plan-vNNN.json. From then on every change is a new plan version, even for two frames.
A changed text or number: change the plan (new version), run steps 5 to 9. Only changed inserts render again.
crewcut render all <plan> runs 2d, 3d and cut in one go when nothing needs a look in between.
Mix first, then the plan
These hold for every plan, before one frame is rendered. The limits come from edit-style.json.
- Symbol before text. Every insert first shows something that happens. A text card (
headline,post) only where the viewer must read a real source: at mosttext_cards_maxsources and at mosttext_share_maxof the full-screen inserts.keywordonly for a number or a name, at mostkeywords_max, each at mostkeyword_max_wordswords. Count by kind before you render; the style check counts again. - Symbolic means 2D and 3D. 3D where an object with an action is the idea. Drawn 2D for processes, amounts, ways and comparisons: shapes that build, move or change. Per spot, pick what a viewer understands in 2 seconds. A scene holds at most one word or one number, on the spoken word.
- The bar is the reference, not "it renders". Compare with
baseline.jsonandbrand/references/. Rented GPUs are for quality: full resolution, enough samples, soft shadows, clean edges. 2D is motion design, not a slide: parts build one after the other, the camera drifts a little, transitions are soft. - Check every scene before the cut. A still frame and a strip of 5 frames next to the reference. If it looks like a slide deck, clip art or a template, throw it away and build it again. Few very good scenes beat many average ones; reuse a scene 2 to 4 times with another word or state.
- Real material, without being asked. A post, an announcement or a third-party result goes into the picture: the post as a card with a mark on the spoken word, a third-party video short on a card with its source visible. Never an invented screenshot, never a company logo you made up.
- Music belongs to the story (
references/sound-and-gpu.md). Instruments only, quiet under the voice, with a source and rights note. - Compute fast, never idle. While machines load, keep working locally (plan, 2D, cut). Every remote step has a time limit. At the end prove that no machine runs.
- Show the list before the big render (step 3): time, kind, one line of what one sees, the number of text cards.
Scenes
2D scenes (HTML, rendered by Chromium, styled by the brand tokens) and 3D scenes (Blender by code, templates/3d/).
Data keys: templates/2d/BLOCKS.md, templates/3d/README.md. Which form for which statement:
references/plan-and-scenes.md, section 3.
| Kind | Scenes |
|---|---|
| 2D numbers and amounts | counter, bars, grid, wave |
| 2D sources and real material | headline, post, clip-card, cards, swap |
| 2D drawn ideas | area, flow, untangle |
| 2D over the speaker (transparent) | keyword |
| 3D | object-word, row, grid3d, stack, slide, box, scale (shared kit templates/3d/stage.py) |
| recording | "kind": "clip" with file, at, speed: a finished recording full screen |
The neutral pack ships few, general 3D scenes. A user's own scenes go to brand/templates/ (never edit the app's
templates). A new scene gets a line in BLOCKS.md or templates/3d/README.md and is not tested until a frame of it
has been looked at.
Mix rules (values from the pack)
- 3D share: about
graphics_mix.3dof the inserts; 2D carries the numbers; the speaker is on screen in between. - Cut on the word, not on the beat. The word in the picture lands 0 to 3 frames before it is spoken.
- Split long speaker stretches with
"speaker": [{"at": s, "zoom": z}](zoom fromzoom_basetozoom_punch, never overzoom_max). - The speaker is never shorter than
speaker_min_secondsbetween two full-screen parts, and at leastspeaker_open_close_secondsat the start and the end. - Sound: music bed at
music_dbwithduck, a hit only at strong spots, typing under counters, a short gap in the music before the most important number. Loudnessloudness_lufs, true peak at mosttrue_peak_db. - Sources honest: a headline card names the real source.
- Never claim a viewing or listening check that did not happen. This session cannot hear: name the times.
When you apply a rule that comes from a named source, you may say so in one line, for example
crewcut say --episode $CREWCUT_EPISODE --cite "fast cutting and keyword captions after Alex Hormozi's editing style" "…".
Rented GPU or local
- 3D frames render in Blender Cycles. Local means the container's CPU: fine for probes at half size and very short scenes. Anything longer goes to a rented GPU.
- Cost first.
crewcut compute estimate --kind blender --frames N [--local-sec-per-frame S]prints time and money for local and rented. Put the estimate on a card (--kind spend, title with the amount), wait withcrewcut card wait <id>. Exit 3 means no: render locally or drop the scene. Only an identity withmay_spend: truerents (the Editor). - Always: still frames at half size first, then the animation.
- Only scene code and fonts go to a rented machine. Never footage, the voice, keys or private data.
- At the end:
crewcut compute destroy --all-ours, thencrewcut compute ledger --episode ID; the report to the user holds the spend and the proof that no machine runs. Also after a failure. - Pitfalls and what the tools do about them:
references/sound-and-gpu.md, section B.
Honest limits
- The agent cannot hear. Every music and voice judgement is a listening list with times for the user.
- Detection of a word rest, a double start or a level jump finds candidates, not certainties.
- The style check measures limits; whether a scene looks good only a person can say. Put the frames next to the reference and say plainly which scenes are still below it.
Planned
- Re-timing 3D renders when word times shift by a few frames, so they are not rendered again.
- Generated music on a rented GPU through
crewcut compute(the studio did it by hand; seereferences/sound-and-gpu.md). - Web page recordings as a render step (Chromium is in the container; no command yet).
- Comparison sheets (scene frame next to the reference) as one command.
- A self-test on the demo episode in
examples/demo/once it exists.
Sources
No connection to or endorsement by these people; we took ideas and wrote our own summary.
- Alex Hormozi (editing and caption style made by his editor Ryan Magin), https://www.youtube.com/@AlexHormozi: fast cutting on the word and short word captions; here reduced to keyword captions.
- fern (YouTube channel), https://www.youtube.com/@fern-tv: inspiration for the 3D clay look only. No values from it.