heditor：创意视频编辑

# heditor — creative video editing (Reels and long form) heditor turns a recording of someone speaking into a short vertical video that holds attention second by second: a strong opening, a visual world built for what is being said, captions synced word by word, sound effects tied to every action, music with an emotional arc and a call to action. Compositions are HTML + GSAP rendered to MP4 by [HyperFrames](https://github.com/heygen-com/hyperframes). Everything runs locally. The standard is high: **every video gets an edit designed for it.** A style is a proven visual language, never a template to stamp. If two videos in the same batch look like the same mould, the work is not done. ## 0. Setup (once per machine) ```bash bash <skill_dir>/scripts/setup.sh # installs ffmpeg/node/python deps if missing, venv, Chrome, fonts, music, SFX, models . ~/.heditor/env.sh && heditor doctor # must end with "heditor is ready." ``` - Re-running setup is safe. `--no-audio` skips the Mixkit library; `--whisper small,medium` pre-downloads ASR models. - Edit `~/.heditor/brand.json` (name, `@handle`, colours, logo path, Follow/Following words, language). Ask the user for these the first time; never invent a handle. - If a tool cannot be installed automatically (no sudo/package manager), tell the user the exact command to run and stop. - All commands below assume `. ~/.heditor/env.sh` was sourced. The CLI is `heditor` (also at `<skill_dir>/scripts/heditor`). ## 1. Choose the mode | User says | Mode | Read | |---|---|---| | "edit in style X", "igual ao estilo X" | Fixed style | `references/styles/<slug>.md` | | "nativo", "edição simples/limpa", "só corta e legenda" | Style `nativo` (the one exception to "different every time": the edit stays invisible) | `references/styles/nativo.md` | | "creative mode", "modo criativo", "surprise me" | Creative mode | `references/creative-mode.md` | | "creative mode with/using X" (GTA 6, Wes Anderson, anime, news broadcast…) | Creative mode + reference | `references/creative-mode.md` (§ Reference direction) | | Nothing specified | Ask once which mode; if the user wants you to decide, use creative mode | | | "vídeo longo", "aula", "live", "YouTube", "16:9", "editar a aula inteira", chapters, SRT | Long-form mode (any style or creative) | `references/long-form.md` | | "plano de gravação", "analise os roteiros", "pré-produção", plan before recording | Pre-production: recording plan, teleprompter and edit plan; nothing is edited yet | `references/pre-production.md` | | "novas aberturas", "teste de gancho", "troca só o começo", new hooks for an ad that already converts | Hook variants: render only the opening and splice it onto the delivered file, body identical (`heditor splice`) | `references/hook-variants.md` | | Raw recording read from a teleprompter or a batch of scripted takes ("pausas para ler", "falei com o editor", "quebrei a quarta parede", retakes, erros de gravação) | Triage first (`heditor triage`): asides to the editor are notes to follow, never content; then any mode | `references/raw-takes.md` | Named styles (`heditor styles`): see `references/styles/INDEX.md` for when each fits. | Style | Identity | |---|---| | `cinema` | One visual metaphor per speech beat, world changes on each turn, live camera card, kinetic captions | | `ruptura` | Cold open with the strongest line, the camera itself becomes the metaphor, typographic finale | | `impacto` | Organised layout with coordinated emphasis bursts (type, sound, zoom) on the key lines | | `gravidade` | The speech has weight: every key noun is a labelled block that lands on its word (baked Rapier physics, `physkit.land_on`); problems pile up, the turn lands as a heavy block that throws the heap, the answer stacks as a straight tower. Reference build: `heditor new <work> <video> --style gravidade` + `beats.json` | | `game` | The screen is a retro RPG: HUD, dialog-box captions, skills unlocked, inventory, CONTINUE? CTA | | `thriller` | Security-camera disaster tape: REC overlay, heartbeat, liquid distortion, blackout, system reboot | | `trailer` | Suspense/launch trailer: hard cuts between the speaker (cool-steel grade, letterbox) and black, dramatic pauses inserted in the speech with silence before every hit, words on the wall behind the speaker (matte), titles spelled by light, a real object of the set as the date device, layered score with a smash to silence before the climax, poster end card. Beats measured on real trailers | | `ide` | Code-editor screencast: circular webcam bubble, captions typed as code comments, terminal CTA | | `aeroporto` | Control tower: radar, split-flap departures board, flight strips, radio-transmission captions | | `chat` | Light messaging app: the speaker's sentences arrive as messages, the bot replies, video-call PiP | | `documentario` | Mini-documentary: letterbox, film grain, serif chapter cards, documentary subtitles, rolling credits | | `documentario-narrado` | Large-scale narrated documentary (alias atlas): the speaker narrates history on a living night globe (lights and routes on the words), generated period paintings in 2.5D parallax, dated cards with hard silence → impact → reveal, narrator as film-treated plate / gold medallion / voice-over, "next chapter" loop | | `podcast` | Talk show / podcast: saturated set, ON AIR sign, voice-reactive waveform, giant karaoke captions | | `missao` | Open-world game homage: sunset neon world, in-world social clip, character intro freeze-frame, mission card, GPS route, stats screen, mission-complete banner, phone-chat CTA | | `tribunal` | Courtroom sketch: the speaker as expert witness on the court monitor, typed testimony captions, exhibits and stamps, a verdict, the jury is the viewer | | `keynote` | Ad · launch stage: presenter cutout in front of a giant LED screen, slides and stage objects per sentence, deadline close-up, "Learn more" slide | | `stories` | Ad · behind the scenes as a sequence of stories: progress bars, tap and cube swipe, hand-made stickers, "Learn more" link sticker | | `predio-24h` | Ad · isometric company model with a day/night cycle: floors light per department, server cutaway, 24 h orbit, second tower for the offer | | `bilheteria` | Ad · retro box office: numbered tickets, the clerk's arched window, a ticket ruler that gains golden extras, golden ticket, turnstile that lets the viewer in | | `porta-de-aco` | Ad · shop front at night: steel shutter descending at each turn, neon signs, "you take" window, closed shutter as the CTA | | `lapidacao` | Ad · luxury jewellery commercial: the tool is closed and the room goes dark, the raw material drops as a rough stone, the cut (laser, facets, flash) reveals the gem with its parts named, a loupe reviews the machine's work, a tray locks the unsaid steps, the offer case slams shut on the warning, sealed date tag, velvet rope CTA, louder score. Reference build: `heditor new <work> <video> --style lapidacao` + `beats.json` | | `manual-de-montagem` | Ad · assembly manual on a blueprint: numbered parts that snap in, same parts in every project, assembly at the turn, kit box | | `carrinho` | Ad · forgotten cart in a shopping app: push notification, items fall into the cart, bonuses as gifts, checkout CTA | | `entrega` | Ad · order tracking: the timeline jumps to "delivered now", a box opens and items fly into the checklist | | `auditorio` | Ad · TV quiz show: each question is a round, A/B answers light green, scoreboard, velvet doors open for the CTA | | `album` | Ad · sticker album: every named item becomes a sticker on its word, collection tray, "missing one" CTA slot | | `cassino-de-precisao` | Authority · noir casino that flips into a precision bench at the turn; the expert as the croupier; engraved rule | | `mesa-de-risco` | Authority · trading desk on a crash day: the resource is an asset, replay, false-safety seal, circuit breaker, green strategy | | `turno-da-madrugada` | Authority · night-shift control room: main monitor, annunciator windows, process line, time-lapse to dawn, shift handover CTA | | `arquivo-cofre` | Authority · archive reading table at night: the expert as a record in a passe-partout, steel drawer, time-stamp clock, microfilm reel that rewinds to the good day, RESTORED stamp | | `aduana-de-dados` | Authority · night border checkpoint between your computer and the internet: exit declaration, X-ray monitor, barrier, regulations underlined, READ THE TERMS stamp | | `laudo-de-sinistro` | Authority · forensic claim report: the expert in the technical photograph, forensic flash, coverage clause filled as said, pen marks TOTAL LOSS → RECOVERABLE | | `circuito` | Didactic · circuit board: the speaker is the main chip, components drawn on their words (DrawSVG outline + icon, decoded silkscreen), copper traces on the verbs, data pulses (MotionPath), icon morphs (MorphSVG), LEDs and flags, silkscreen titles (SplitText), push-button CTA. Reference build: `--style circuito` + `beats.json` | | `lousa-de-aula` | Didactic · interactive chalkboard lesson: every stroke drawn live on its word, syllabus whose "?" turns into ✓, arrows labelled with the spoken verb, eraser, SUMMARY chain, homework CTA | | `nativo` | Native · clean edit: only cuts at clean word boundaries, jump cuts that alternate the framing, and animated word-by-word captions; no music, SFX, graphics or end card. Reference build: `heditor new <work> <video> --style nativo`, `heditor produce <work> none 0` | | `camadas` | Layers · every concept drawn between the speaker and the wall: giant words and "?" behind the head, guide lines on the set, glass slab, floating cards, comment-to-DM CTA | | `raio-x` | Layers · the machine's view: live pipeline HUD, depth scan, exploded layers, point cloud, face/hand tracking, transcript becoming the prompt, timeline assembly | | `duelo` | Creator · A/B comparison page: the same take in two synchronised panels edited two ways, VS badge, checklist verdict, vote CTA | | `bancada` | Creator · proof on the workbench: label sticker hook, whip pan to a filmed/simulated monitor, finger and punch-ins on the screen, result inserts | | `manchete` | Creator · headline inset on top: subject clip/post above, banner across the seam, the speaker reacting below, full-bleed peak, band splits | | `reacao` | Layers · the speaker in front of the post: raw talking head cut out at the bottom, the news/post/page/screen recording full-bleed behind him, pushed in, scrolled and panned on his words, boxes, highlighter, underline and circles on the exact word, lime highlighter headline, title cards, credit chip, profile CTA. Backdrops: screenshot, `heditor capture <url>`, `heditor region` (the post inside the raw recording) or a clip. Reference build: `heditor new <work> <video> --style reacao` + `beats.json` | | `revista` | Creator · editorial spreads: condensed keyword + italic serif line, rounded portrait window, icon chips per item, date pages, full-frame breathers | ### Long-form mode (classes, lives, YouTube; any format) For minutes to hours of video in 16:9 (also 16:9 4K, 1:1, 4:5, 9:16), follow `references/long-form.md`: `heditor new --long <work> <video> --format 16:9 --style <slug|creative> --kind aula|live|youtube` → `heditor long transcribe` (chunked, resumable ASR) → `long edl` (proposed cuts of dead pauses, false starts, repetitions and fillers, reviewed before applying) → `long apply` → `long chapters` → `long layouts` (screen + camera layouts: pip, screen with zoom, speaker, side) → `long produce` (style pack, set pieces, captions, lower thirds, intro/CTA, mix) → `long preview` → `heditor test` → `long render` (segmented, resumable) → `long finalize` (one continuous master audio, SRT/VTT, YouTube chapters, QA report). Every named style is available as a *style pack* (`heditor long styles`); creative mode lives in set pieces at key moments. Format profiles: `heditor formats`. ### Layered effects (any mode) Text and objects *behind* the speaker, captions painted on a wall, 3D objects passing between the speaker and the room, depth fog, contour scans, point clouds and objects that follow the hands come from depthkit: `heditor matte|depth|look|hands|plane <work>` (models download on first use). Read `references/depth-and-3d.md` before using them in any style or in creative mode; layouts, illustrations, hooks and CTAs borrowed from creator Reels are in `references/technique-catalogue.md`. ### Motion tools (any mode) - **GSAP plugins offline:** SplitText, DrawSVG, MorphSVG, MotionPath, CustomEase/Bounce/Wiggle, Text, Physics2D/Props and EasePack ship with setup; load them with `rk.page(..., plugins=("SplitText", "DrawSVG"))`. Read `references/gsap-craft.md` first: some GSAP features call `Math.random` (ScrambleText, `stagger from: "random"`, `gsap.utils.random`) and break the render; the kit has seeded replacements (`rnd()`, `scramble()`). - **SVG preparation:** `heditor svg <in.svg> [out.svg] [static|draw|morph]` cleans exported SVGs with SVGO for the animation they will get and proves, pixel by pixel in the render Chrome, that they still look the same. - **Real physics:** `heditor physics` / `physkit` bakes a Rapier rigid-body simulation (falling piles, impacts, blasts) into per-frame transforms that a three.js layer with bloom and depth of field reads at any time. Recipe, framing and look values in `references/physics-3d.md`; working pilot in `templates/physics/pilot_build.py`. - **Styles built on these tools:** `gravidade` (physics: blocks land on their words, `physkit.land_on`) and `circuito` (DrawSVG, MorphSVG, MotionPath, SplitText on a circuit board). Both ship a reference build: `heditor new <work> <video> --style gravidade|circuito`, write `<work>/beats.json` from the speech (format in the style file and at the top of the build), then `heditor produce`. In creative mode, borrow their engines (`board()`, `physics()`) for a single set piece. ### Stock video / B-roll (any mode) `heditor stock search "<english words>" [--orientation portrait]` lists free Mixkit clips (no account, no API key) and `heditor stock get <work> '#N'` downloads one into `<work>/composition/assets/stock/`, after confirming the Mixkit Stock Video Free License on the clip page, and records its provenance in `<work>/stock.json` (paste the printed line into DESIGN.md). Use B-roll only where it illustrates a word of the speech, muted, 1.5–3 s per idea, without recognisable faces next to claims in ads, logos or brands; one clip per `get`, never bulk downloads (Mixkit's terms forbid scraping the library). Read `references/stock-media.md`. ## 2. Workflow (all modes) Follow `references/workflow.md` step by step. Summary: 1. **Probe the source.** `heditor probe <video> 5 60 300` and look at the frames: where is the speaker (full frame, webcam tile in a meeting, screen share)? Note resolution, the speaker crop, what else is on screen. 2. **Transcribe and choose the cut.** `heditor transcribe <video> 0 <duration> qa/asr-full.json` (or a window). Pick a self-contained argument of 30–70 s: hook → thesis → explanation → conclusion. Long-form stories (e.g. a narrated documentary of ~90 s, style `documentario-narrado`) are allowed when requested: declare `max_duration` in the plan (≤ 180 s) so `heditor test` accepts them, and keep a retention device every few seconds. The hook must exist in the speech. Never invent, reorder words inside a sentence, speed up or splice half-sentences into a new claim. Remove repetitions, false starts and tangents only at clean word boundaries. Scripted or teleprompter takes go through `heditor triage <video...>` first: chunks between silences transcribed alone and labelled keep / retake / false_start / editor / noise, so reading pauses, repeated sentences and asides to the editor never reach the cut. An aside is a note from the speaker to you (`cut`, `redo`, `marker`, `note`): follow it and report what you did (`references/raw-takes.md`). 3. **Scaffold and prepare.** `heditor new <work> <video>`, fill `BLOCKS`/`CROP` in `<work>/prepare.py`, `heditor prepare <work>`, `heditor gaps <work>` (no pause > 0.45 s), then `heditor words <work>` and fix spelling/brand names in `words-reviewed.json` (timings untouched; profanity can be masked on screen only). 4. **Design first.** Fill `<work>/DESIGN.md`: video analysis, mode/style, frame map, palette, captions, beat sheet with timestamps, music, CTA, risks. In creative mode, research before designing. 5. **Build.** Rewrite `<work>/build.py` (the template is only a runnable baseline) to implement DESIGN.md. Use the kit (`references/kit-api.md`) or write your own HTML when the format needs it. 6. **Produce.** Pick the mix profile (`voice` or `presence`), profile the track with `heditor music <music:ID|file>` and place its turns on the speech turns; `heditor produce <work> music:<id> <drop_s>` → composition + mix; read the `balance` it prints (targets, `music_gaps`). 7. **QA before render.** `heditor lint <work>` (0 errors) → `heditor preview <work> <18–24 times>` → inspect the contact sheet with vision and fix every defect → `heditor test <work>`. Gates in `references/quality-gates.md`. 8. **Render and finalize.** `heditor render <work>` → `heditor finalize <work> <name>.mp4` → review `qa/final-contact-sheet.jpg` and `qa/final-verification.json` → `heditor asr-check <work>`. 9. **Deliver** the MP4 path with a short summary (format, music, what was verified, what nobody has checked: nobody has *listened* to it — say so). Never publish to social networks unless the user explicitly asks. 10. **Learn.** If you created a new style that the user liked, add it as a named style (`references/adding-styles.md`). ## 3. Non-negotiables - **No dead air.** Inspect pauses inside blocks and across every junction, using the assembled clean voice, not the music mix. Run `heditor gaps` after prepare and check the surrounding words. The default 0.45 s threshold flags candidates for revision; it does not authorize clipping consonants, removing every breath or flattening deliberate rhetorical pauses. Preserve speech at 1× and shorten unwanted dead air with small safety margins. Document any intentional longer pause and verify the exception explicitly before rendering. For a pause-only revision of an approved edit, preserve its visual language and previous version, remap captions, media and SFX to the new timeline, then verify the exported MP4. - **Voice first, music present.** Mix profile by delivery (`plan["mix"]`): `voice` (default; organic Reels, class excerpts) music ≈ 21.5 dB under the voice (median), background ≥ 12 dB under at p95; `presence` (paid ads, creative mode, trailer) music ≈ 15 dB under, background p95 ≥ 7 dB under, filtered sections with make-up gain and `heditor asr-check` required. In both the bed may not vanish inside the speech (`test_music_presence`). Impacts in the gaps, ducking held across the whole phrase. -14 LUFS, true peak ≤ -1.5 dBTP after AAC. - **Readable captions.** Every spoken word, synced; active word with high contrast (black on light yellow, white on saturated dark — never dark-on-mid or dark outline on dark text). Check enlarged crops. - **Safe zones.** (9:16; every format profile has its own safe area, see `long-form.md`.) Nothing important above y = 270 (top 14%; titles with accents start ≥ 300), keep the bottom ~350 px and the right edge (~120 px, y 1000–1700) free of critical info, never cover the face.