edit-video：口播原片直出 YouTube 成片

edit-video — talking head → final-16x9.mp4 + final-9x16.mp4

You are the editor. Tools in tools/ do the mechanical work; you make every editorial decision and keep it in files (work/edl.json, work/visual-plan.json, work/audio-plan.json) so any stage can be redone. Read docs/lessons.md once per session — every rule there was paid for with a broken render.

Architecture (why it is shaped like this)

raw ─ transcribe ─ cut ─┬─ work/cleancut.mp4 ──────────────┐
                        │                                    │ (dialogue)
       visual-plan.json ┴─ Remotion `Doc` (muted) ×2 formats  │
                             └─ HyperFrames captions pass ×2  │
       audio-plan.json ─────────────── tools/mix.mjs ◄───────┘ → work/mix.wav (once)
                                                     finalize: picture + -14 LUFS master + intro/outro

• Remotion draws the whole picture: the a-roll in one of six stage layouts (full / dock left / dock right / band / corner / takeover), 28 motion-graphic types, b-roll cutaways, camera moves, chapter rail.
• HyperFrames draws only the word-level captions (browser shaping renders Telugu/Devanagari correctly).
• ffmpeg does ALL audio: no <audio> in any render — embedding it corrupted dialogue repeatedly in the template this repo inherits from.
• Both formats share one clean-cut timeline, one visual plan and one mix.

Stages

#	Stage	Tool	Skill / notes
0	Intake	npm run new -- <inbox-name> --brief "…" [--lang te|hi|en] [--style glass|broadcast] [--keyterms "A,B"]	see Intake below
1	Transcribe	npm run transcribe -- <p>	English → whisper (free) · Telugu/Hindi → ElevenLabs
2	Clean cut	npm run cut -- <p> [--dry-run]	clean-cut skill. Most important stage.
3	Visual plan	write work/visual-plan.json; npm run visuals -- <p> --check; --stills t1,t2	motion-graphics skill
4	B-roll	npm run broll -- search|get	inside motion-graphics
5	Render picture	npm run visuals -- <p>	~1-2 min per minute of video per format
6	Captions	npm run scaffold -- <p> (+ npm run captions -- <p> --style …)	captions skill
7	Audio plan	write work/audio-plan.json; npm run sfx -- list; npm run bgm -- search|get; npm run mix -- <p>	sfx, bgm skills
8	Finalize	npm run finalize -- <p>	lint, render, master, intro/outro, output/credits.md
9	Verify	npm run verify -- <p> → Read both work/verify-*/sheet.jpg	never skip
10	Package	youtube-metadata, youtube-thumbnail skills	offer at the end

npm run status shows progress. After a re-cut: node tools/retime.mjs <p> carries both plans onto the new timeline (old clean → raw → new clean) and lists anything whose moment was cut.

Running it

1. Intake. The user drops the clip in inbox/talking/ and names it in the prompt ("edit video-1 …", possibly written "/video-1" — strip the slash). npm run inbox lists what is waiting. npm run new -- video-1 --brief "<the user's style words, verbatim>" (+ --lang if they said it, --keyterms for names/brands). The prompt's style instructions are the brief — map them to settings before planning (see "Style brief → settings"). Ask nothing the prompt already answers. Brand/channel: none by default — this repo edits for many clients. Only when the user says whose video it is, load config/brands/<slug>.json (or create it from what they tell you) and pass --brand, plus --intro/--outro clips from inbox/brand/ if they ask for them.
2. Transcribe, then read work/transcript.md end to end. Check the header's engine and language.
3. Clean cut — follow clean-cut exactly, including the boundary-audit and fragile-word output.
4. Plan visuals from work/cleancut.transcript.md (clean-cut times). Write the plan, --check it (it validates every component's required fields — the template's #1 silent defect), fix, then render --stills at 6-10 interesting times and Read them in both formats before the full render. Framing: if the speaker is off-centre in 9:16, set "framing": {"9:16": "40% 45%"}.
5. Render picture, scaffold captions, plan audio and mix, finalize, verify.
6. Report: both output paths; raw → clean duration and what was cut (by reason); overlay/b-roll/SFX counts; music used; anything fragile or uncertain; credits that MUST go in the description; and say plainly that you cannot hear the mix — ask the user to listen for the joins, bed level and ducking. Offer /youtube-metadata and /youtube-thumbnail.

Style brief → settings

Translate the user's words into concrete choices and state them in one line before planning:

They say	Set
fast / energetic / reels / punchy	keepPauseMax 0.25, caption pop 2-3 words, a beat every 2-4s, camera punches, sub-band SFX on structure
calm / professional / documentary / long-form	keepPauseMax 0.4-0.5, caption clean 4-6 words, a beat every 6-10s, fewer SFX, soft bed
minimal / clean / no-fuss	few overlays (text marks only where useful), no marquee/glitch, clean captions or none
bold / premium / news / editorial	--style broadcast, high-contrast accents
storytelling / emotional	karaoke captions, piano/cinematic bed, a music gap on the key line
"no music" / "no SFX" / "no captions" / "no b-roll"	honour exactly; skip that stage
(music not mentioned)	no BGM — the user says at the start of the prompt whether to add music; default is none. When added: CC0 only (--cc0) and SFX must stay clearly audible over it
a colour ("use yellow", a hex)	caption --accent, overlay section accents/brand

Anything unusual in the brief overrides these defaults. Keep the brief in project.json.brief so a re-edit later follows the same style.

Review gates

Default: run straight through and report. If the user asks for checkpoints (or project.json.review is "stages"), stop after the clean cut, after the visual plan (show stills), and after the audio plan. A requested change is the new bar for that stage — revise and re-show the same stage.

Rules

• All plan times are on the clean-cut timeline; EDL times are raw.
• Overlay text is short English by default (native script lives in the captions) unless the brief says otherwise.
• Never hand-edit generated files (compositions/captions.html, work/props-*.json) — regenerate.
• Keys in .env: ELEVENLABS_API_KEY (Telugu/Hindi), PEXELS_API_KEY or PIXABAY_API_KEY (b-roll). Missing one → stop and say exactly which, and that it goes in C:\AI-edits\.env.
• Windows: write JSON with the Write tool (PowerShell mangles inline JSON; Out-File/Set-Content add a BOM — the tools tolerate it in JSON, but never write code files that way).
• Cleanup once the user confirms the final: render/, work/visuals-*.mp4, work/captioned-*, work/stills-*, work/verify-* are regenerable. Keep output/, work/*.json, assets/.