@autoaiai/aicut 0.1.11 → 0.1.13

package/skills/ghost-editor/docs/PLAYBOOK.md

@@ -1,139 +0,0 @@
-# Playbook: every style, every building block, everything we learned
-
-The one-page inventory of this skill. Each style says how to ask for it.
-
-## Named styles (recipes)
-
-| style | ask for it | what it is | best for |
-|---|---|---|---|
-| **Clean OS** | "edit this in Clean OS" | An Apple-ad talking head that lives inside your phone. 1–3 lowercase word cards resolving from blur (Inter), wide/tight cuts from one 4K camera, literal iOS scenes for what you say (date picker, lock-screen notifications, a search with a loading bar stuck at 97%, a Photos memory card), a red Figma selection box with a cursor, perspective text with a draining ring, one handwritten word (Great Vibes), a small sound on every graphic. | reflective or explanatory monologues |
-| **Vector explainer** | "make an animated explainer" | No footage. One continuous SVG world a camera travels through, a subject that draws itself, serif headlines on panels, words that rise from blur as the narrator says them, an ElevenLabs narrator, a piano bed. | parables, life lessons, concepts you can draw |
-| **RESULT** | "edit this in RESULT style" | A warm podcast edit: one or two Inter Tight words with a small line and one big word, Caveat handwriting, navy cards with film dust, polaroids, a paper stack that swaps pictures stop-motion, film-burn light leaks, comic props (REC frame, an incoming call you decline, a cursor clicking a dialog), dissolves between lines. | storytelling, nostalgia, podcast monologues |
-| **POV montage** | "cut these POV clips into a reel" | Several first-person clips cut into a story: a flash-forward hook, speed-ups, text behind you, a cursor-click fact card, a freeze frame with a zoom, a record scratch and a meme, on the editorial motion layer. | trips, vlogs, a day from smart glasses or a GoPro |
-| **Seoul list** | "Seoul list", "a list reel over B-roll" | A voice-over list over fast B-roll (1.4 s shots, no talking head): one big word per line, italic serif (Playfair) for places and moods, heavy Inter Tight for facts, with a subline tucked under it; white word flashes; white collage pages in the hook (the shot shrinks into a card, a red glitch word, a photo card flying in, Fluent hand and eye stickers); triptychs; an Instagram comment card for the personal aside; a macOS frosted window swapping cut-outs; blue highlight words; an iMessage outro (typing bubble, word bubble, emoji tapback). | "N things about a city / a trip / a topic" read over B-roll: yours first when you hand it over, else your library, else stock or AI in one graded world |
-| **stan style** | "stan style", "tool tutorial" | A seated talking head explaining a tool, with real screen recordings popping into the top band behind the speaker (the head is cut out and overlaps the screen; the speaker moves down only if the eyes would hide), small white lowercase chest captions that cut phrase by phrase (single words in the hook and the diagram, forehead words near the end), a hook card (icon rows + a black-on-white strip title), one white diagram whose curved lines draw on after a zoom-blur, «[Step n]» labels with app icons under the chin, a comment-keyword CTA with a logo, «*asides*» above the head. | "how to do X with tool Y" in steps, with screen recordings |
-
-Each lives in `recipes/<name>/` with a README (the full grammar and pipeline), a SYSTEM.md, its scripts and a component catalog; the vector explainer, Seoul list and stan style also ship an English worked example.
-
-## Presets (one line in reel.json: `"style": "<name>"`)
-
-| style | looks like | best for | sound · music |
-|---|---|---|---|
-| `clean` | bold mixed-case captions, one keyword colour, snap zooms | business advice | restrained · none |
-| `editorial` | lowercase blur-in captions, tilted pill tags, highlight wipes, full-screen scenes | data, insight | standard · tense electronic |
-| `meme` | outline captions, emoji and memes beside the head, chips and big cards | entertainment, hot takes | rich · quirky |
-| `cinematic` | film grade, grain, vignette, B&W B-roll, glowing serif keywords | stories, founder journeys | restrained · ambient |
-| `launch` | dark UI windows with typed prompts, phone chats, glitch cuts | product and AI launches | standard · upbeat electronic |
-| `kinetic` | full-screen word-by-word type, speaker in a round window | hooks, quotes, manifestos | standard · percussive |
-| `pop` | giant words behind the speaker, uppercase captions in hot colour boxes | personal brand, bold claims | standard · upbeat pop |
-
-Or show it a reel you love: `scripts/reference_study.py` reverse-engineers it (Gemini 2.5 Pro edit log + frames) and the agent maps it onto your recording. A look you like can be saved as a new preset or recipe.
-
-## Pipeline
-
-0. `doctor.sh`: tools, keys, library, each with its fix.
-1. `prep.sh`: 1080×1920, voice at −16 LUFS; a contact sheet; confirm the right person is on screen.
-2. `denoise.sh` for outdoor takes (DeepFilterNet3, capped at 20 dB).
-3. `transcribe.py` (every take, word times) and `face_track.py` (keeps captions off the face).
-4. Choose the best complete take per line (or `autocut.py` for one clean take); re-transcribe the cut; Gemini 2.5 Pro verbatim audit.
-5. Pick a preset or recipe; plan beats on spoken words.
-6. `build.mjs` (taste defaults + warnings), then `approval_sheet.py`: one page with a frame at every cut and beat, the timeline and every warning; the user approves before the render.
-7. Render; `render_report.py`: pass/fail per check with a fix each, at most 3 fix-and-re-render rounds; `qa.py` for SFX against the voice on a stem, A/V lengths, matte, transitions and shader flicker.
-8. Deliver; `taste.py log-reel`; `save_style.py` keeps a look the user loved as a preset.
-
-## Building blocks
-
-- **Captions:** house (Arial bold 56, outline, lower third), pill, clean, editorial (lowercase blur-in, tags, highlight wipes), mini, word cards (Clean OS).
-- **Caption templates (v0.3):** hype, explosive, line-focus, neo-minimal, karaoke-fill, highlighter, with one keyword per phrase (heuristic, forced or Gemini).
-- **Kinetic type (v0.3, `textfx`):** hook (a giant cropped word at frame 0), drum (3D word cylinder), scramble (glyphs lock on the word), flipstack (cards pile up), headline (masked rise + brush mark, "not X -> Y"), spring (letters on a spring).
-- **Shader looks (v0.3, Paper Shaders):** warm-grain, mesh-calm, ink-warp, god-rays, dither-retro, fluted-self (ribbed glass over your own frame) behind a word scene; film-grain and halftone over the speaker.
-- **Animated emoji (v0.3):** 50 Fluent reactions (40 animated) that pop next to the head with overshoot.
-- **Transitions (v0.3):** 10 eased ones on stock ffmpeg (dissolve, zoom blur, soft wipe, crosswarp, warp wipe, dip to black, whip, scaled push, flash, static wipe) laid over the cut; the voice stays a hard cut.
-- **Film looks (v0.3):** six film stocks (Portra 400, Vision3 500T, Eterna 500, Ektachrome 100D, Tri-X 400, CineStill 800T) with datasheet grain and halation, baked into the plate.
-- **Person matte (v0.3):** BiRefNet cut-out for words behind you, computed only for those seconds.
-- **Cutout effects (v0.3):** one BiRefNet pass per 2-3 s beat: background swap, portrait blur, selective colour, spotlight, die-cut sticker freeze, neon outline, clones, pop out of a phone frame; reel-wide selective/spotlight/blur looks (`look.cutout`); covers for Reels and YouTube (`cover`). Words can also slide through the speaker (`behind` + `through`).
-- **Card beats on the word:** big (claim slam, count-up), chips, strike, quote, list + stamp, emoji / logo / meme (reaction slot), behind (giant word behind you), title, broll, endcard.
-- **Full-screen scenes:** card (pill + cursor click), stats (glass badges), fly3d (camera through 3D words), image (still B-roll), ui (typed prompt), device (phone chat), kinetic (word-by-word + round window), sentence (hero word lands).
-- **Motion:** hard cuts by default; blur / expand / wipe / glitch only at section changes; snap zooms; slow pushes; wide ↔ tight; freeze frames.
-- **Layout:** face-aware placement (below the chin, above the head, shrink; never on the eyes or mouth) inside the Instagram / TikTok / Shorts safe area, clear of the like column; one card and one reaction at a time.
-- **Look:** grade, grain, vignette per style; person matte for text behind you; keep the footage's own mood.
-
-## Sound
-
-- **Voice tools (v0.3, optional):** ClearerVoice super-resolution for band-limited mics, silero-vad speech islands (`autocut.py --vad`), a picture-vs-voice length check in qa.py.
-- **Effects:** UI (pop, click, tick, keyboard, notification, ding, shutter, a message-send swoosh), whooshes and risers, impacts and booms. Only Mixkit sounds (free licence), fetched at setup; add your own memes and meme sounds with `meme_add.py`.
-- **Music beds (13, Mixkit):** calm (lo-fi-02, digital-clouds, silent-descent, vastness), tech (deep-techno-ambience, cyberpunk-city, infinity, brainiac), fun (comical, keeping-fit, happy-times, life-is-a-dream), rhythm (vocal-percussions).
-- **Mixing:** the voice is levelled first; each SFX role sits a fixed dB under your measured voice peak; music ducks under speech (`music.duck`); final −14 to −16 LUFS, true peak ≤ −1 dB.
-- **Voices:** ElevenLabs (`tts_elevenlabs.py`, e.g. Spuds Oxley for a wise elder), Kokoro as a free local fallback.
-
-## Type (free, OFL, Latin + Cyrillic)
-
-Inter (Clean OS, stan style captions and labels), Inter Tight (RESULT, Seoul list, stan style titles), Great Vibes (script word), Montserrat (presets), Playfair Display (cinematic, explainer headlines, Seoul list's italic serif), Geist and Geist Mono, JetBrains Mono (UI windows), Unbounded, Oswald. Arial Bold for house subtitles.
-
-## What we learned (each rule came from a real mistake or a measurement)
-
-**Cutting and takes**
-- Transcribe everything and choose takes from the table; cut on word times, never segment times.
-- Whisper merges restarts into one clean sentence and smears timestamps. Re-transcribe the cut voice and audit it with Gemini 2.5 Pro.
-- A long stretch of voice with few transcript words is a hidden retake. A throat-clear is a short voiced hum with falling pitch (find it with a pitch scan).
-- Leave ~0.2 s of breath between sentences. Confirm the right person on screen before rendering.
-- Build an attempt log of the whole raw take with Gemini 2.5 Pro before cutting. Hide jump cuts with short dissolves, not a zoom on every cut (three comic punch-ins at most).
-
-**Audio**
-- Full-strength denoise leaves a thin, watery voice: cap at 20 dB and add a little warmth.
-- Level the voice first; never add gain to the finished mix. The SFX check must exclude the music bed.
-
-**Captions and layout**
-- Never cover the face; stay inside the safe area. Light accents need dark text. Hide captions while a headline says the same words. No em dashes on screen; one emphasis mechanism per video.
-
-**Motion and rendering (HyperFrames)**
-- Anything tweened twice needs `fromTo` with explicit values; hide each caption with a `fromTo` placed after its fade-in. Grow SVG shapes with `attr`, not scale. Every video needs an id. Lint, snapshot every beat, look, then render.
-- Top reels: 113 of 119 transitions are hard cuts, median shot ~3 s, B-roll 20–35 % of runtime, audible SFX 0.4–0.9 per 10 s on graphics, not cuts.
-
-**Sound and memes**
-- Measure a reference's sound before copying it: the reference Seoul reel has no audible bed and 3 non-voice hits in 46 s (mix vs a
-  speech-enhanced copy), where Gemini's edit log claimed a lo-fi bed and ~25 hits.
-- A small sound on every graphic, none on bare cuts. Never the same meme sound twice; 1–3 per minute; never on a word. Meme clips from the internet are other people's media: Instagram and TikTok at your own risk, never YouTube (Content ID).
-
-**Footage and B-roll**
-- 4K vertical gives two shot sizes from one camera. No 720p B-roll; on Mixkit only clips offered in 2160p are free for commercial use. Your own unused frames make good B-roll.
-- B-roll is the user's world, framed and graded, or nothing (references/broll.md, from an audit of 29 reels).
-  - Sources in order: the user's own footage, then stills in one frame language (AI or stock images), then stock video
-    that matches the reel's world.
-  - Hard rules: the configurable B-roll people rules (`people-rules.json`, empty by default; e.g. exclude women,
-    children or symbols), checked by Gemini on every shot. Every shot shows its line. The speaker's face is never regraded.
-  - One character per idea. A spoken list becomes typography.
-  - Talking heads: <= 3 inserts per minute, <= 25 % full-screen, the speaker's face on screen >= 65 %.
-  - Every insert is graded onto the speaker's plate as shot (`insert_prep.py`); `broll_vision.py picks` shows the user every shot first. `broll_qa.py` fails the render past these limits.
-- Small glances stay on screen. Cover only real look-aways (> 0.5 s or > 12 deg); hiding every flagged frame put one
-  reel's speaker on screen 33 % of the reel.
-- Defaults:
-  - stock only as a last resort and inside a frame (talking-head styles; Seoul list is full-screen B-roll by format);
-  - the user's own footage found automatically with a face check (only they appear), and the picked frames shown to
-    them before the render;
-  - one fixed house look per style (`recipes/<style>/house.json`): type, captions, accents, insert framing, the
-    speaker's face as shot (never regraded).
-- New components every reel in a named style: scenes, props, polaroid / collage treatments, transitions, hooks and
-  outros not used in the style's last 7 reels (`components.py plan / check / record`).
-- stan style needs real screen recordings; without them, propose another style rather than filling the band with stock.
-  Seoul list is a format, not a footage source (the footage the user hands over first, then their library, then stock
-  or AI full screen), held to one graded world (conformed to the main footage as shot) with no posed stock models.
-
-**Workflow**
-- Nothing renders before the user approves the approval sheet; a failed render report gets at most three fix rounds, then the user decides.
-- One reel at a time: no batch render of a wave of styles without a sheet per reel (7 styles from one template in 7 minutes came out with almost no approval files).
-- Every correction is recorded as taste in the user's words; learned rules are tagged as the skill's, never as "you said".
-- Cut picture and voice at the same frame-snapped times, or the voice drifts early across the pieces.
-
-**Tools**
-- Pin the HyperFrames CLI (`npx -y hyperframes@<version>`): an unpinned `npx hyperframes` fetches whatever was just published, and a release whose tarball 404s fails every render mid-batch.
-- Phone-filmed screens show the browser chrome (tabs, a URL with access tokens, bookmarks). Find the lowest line the chrome ever reaches across the cut, and clamp every zoom so nothing above it is ever in frame.
-- Snapshot every beat and read the strips before rendering (recipes/seoul-list/snap.py): the first Seoul-list build had
-  six visual bugs (ghost copies visible from frame 0, collapsed word gaps, an empty card after its footage ended) that
-  lint and the timeline passed.
-- Check a reference's geometry on full-res frames, not just its look: the stan-style breakdown said the speaker is "reframed
-  down" under the panel, but face tracking on the reference put the eyes at y 815 in both split and full shots; and the
-  panel is BEHIND the head (the cut-out hair overlaps the screen: it was first read the other way round and only the
-  smoke render showed it); its title was black on white, not white on black; the "24/7 strobe" is not in the frames.
-- Kinetic exits that stagger per letter or card must finish by the block end (fixed in kinetic.mjs: `t >= t1` forces the exit); also give each block a hard visibility window in the page.
-- Search the HyperFrames registry (`npx hyperframes catalog --query`) before hand-building a look. Validate API keys with one call before a batch.

package/skills/ghost-editor/docs/THIRD_PARTY/audio.md

@@ -1,17 +0,0 @@
-# Sound effects and music (Mixkit)
-
-No audio file ships in this repo. `library/sfx/manifest.json` and `library/music/manifest.json` list Mixkit items by id
-and source page; `scripts/library_restore.py` and `scripts/sfx_fetch.py` download them onto your machine.
-
-- **Sound effects:** Mixkit Sound Effects Free License. Use them in your videos, commercial included, no attribution.
-  You may not redistribute an item on its own, as stock, or inside a tool or template, which is why only the list ships.
-- **Music:** Mixkit Stock Music Free License. Fine for web and social video, podcasts, ads and YouTube. Not allowed:
-  CDs/DVDs, video games, TV or radio, remixing into a music-only track, or registering a track with Content ID or any
-  other rights service.
-- **The download script:** Mixkit's User Terms (mixkit.co/terms/) say not to "use scripts or bots to mass download
-  Items". The restore fetches about 70 items, once, from your own machine. If you'd rather not, download the items by
-  hand from the `source` pages in the manifests and save them under the manifest's `file` names
-  (`library/sfx/<id>.wav`, `library/music/<file>`); `doctor.sh` checks the files, not how they arrived.
-- Full terms: https://mixkit.co/license/
-
-Memes are not included. Clips and sounds you add with `scripts/meme_add.py` are your responsibility.

package/skills/ghost-editor/docs/THIRD_PARTY/audioqa.md

@@ -1,13 +0,0 @@
-# Third-party code and weights: audio QA
-
-Venv `~/.venvs/ghost-audioqa` and weights in `~/.cache/ghost-audioqa`, both made by `scripts/setup_audioqa.sh`; nothing
-below is committed.
-
-| What | Source | Licence | How it is used |
-|---|---|---|---|
-| ClearerVoice-Studio (`clearvoice` 0.1.2) | https://github.com/modelscope/ClearerVoice-Studio | Apache-2.0 | Library imported by `scripts/voice_enhance.py`. At run time `clearvoice.utils.decode.bandwidth_sub` is swapped for our own band merge (`detect_cutoff` + `merge_bands`). |
-| MossFormer2_SE_48K / MossFormer2_SR_48K weights | https://huggingface.co/alibabasglab | Apache-2.0 (model cards) | Downloaded at setup (`voice_enhance.py --fetch-only`), inference files only (the SR repo's 1.7 GB training checkpoint is skipped). |
-| silero-vad 6.2.3 (weights inside the package) | https://github.com/snakers4/silero-vad | MIT | Imported by `scripts/vad_pauses.py`. |
-| pyloudnorm, torch/torchaudio/torchvision, numpy, opencv-python, soundfile, matplotlib | PyPI | MIT / BSD / Apache-2.0 / PSF-style | venv dependencies. |
-
-No GPL/AGPL code was used.

package/skills/ghost-editor/docs/THIRD_PARTY/captions.md

@@ -1,20 +0,0 @@
-# Third-party sources: caption templates
-
-Files: `scripts/lib/caption-templates.mjs`, `scripts/lib/ai-tag.mjs`, `scripts/lib/metrics.json`, `scripts/font_metrics.py`.
-
-No third-party code was copied. Everything was written for this skill; these sources were studied for ideas only:
-
-| source | licence | what we took (ideas only) |
-|---|---|---|
-| pycaps, https://github.com/francozanardi/pycaps (v0.3.0) | MIT | The template names and looks `hype`, `explosive`, `line-focus`, `neo-minimal`. A template as style + splitter + animations + tagger rules. Per-word narration states. Structure tags (first/last word in a line or segment) and semantic tags exposed as CSS classes. The `limit_by_chars` splitter with "don't end a segment on a short word". |
-| remotion-captions-kit, https://github.com/Fats403/remotion-captions-kit | MIT | Karaoke fill as a `clip-path: inset()` sweep driven by word progress. Pagination: break on silence, sentence ends and a character cap; avoid one-word orphans. |
-| HyperFrames registry components `caption-highlight`, `caption-pill-karaoke`, `inline-highlight` (hyperframes 0.8.146) | Apache-2.0 | How a registry caption stays seek-safe: one group per time window, per-word background sweeps. |
-| HyperFrames skill rule `css-marker-patterns.md` | Apache-2.0 | Highlighter-marker mechanics (an accent box sweeping behind the text); we use a hand-drawn SVG band revealed by `clip-path` plus a dark-ink text layer. |
-| captioncat, https://github.com/ItisShikhar/captioncat | MIT | Skimmed for preset ideas; nothing used. |
-
-Data: `metrics.json` holds glyph advance widths measured from the bundled OFL fonts (Inter, Montserrat, Unbounded, Oswald,
-JetBrains Mono, Geist; SIL OFL 1.1) by `scripts/font_metrics.py` with fontTools (MIT) + brotli (MIT) in the optional venv
-`~/.venvs/ghost-captions` (`scripts/setup_captions.sh`). Numbers only, no glyph outlines.
-
-Service: `ai-tag.mjs` calls Google Gemini (`gemini-2.5-flash`) with `GEMINI_API_KEY`; optional, the default keyword picker is
-an offline heuristic.

package/skills/ghost-editor/docs/THIRD_PARTY/fonts.md

@@ -1,17 +0,0 @@
-# Fonts (`templates/fonts/`)
-
-Every bundled font is under the SIL Open Font License 1.1; each family's licence text sits next to it as `LICENSE-<family>.txt`.
-
-| family | files | source | reserved font name |
-|---|---|---|---|
-| Geist, Geist Mono | `geist-latin`, `geist-mono-latin` | Google Fonts (Vercel's Geist) | none |
-| Inter | `inter-latin`, `inter-cyrillic` | Google Fonts | none |
-| Inter Tight | `inter-tight-*` (4, incl. italic) | Google Fonts | none |
-| JetBrains Mono | `jetbrains-mono-*` | Google Fonts | none |
-| Montserrat | `montserrat-*` | Google Fonts | none |
-| Oswald | `oswald-*` | Google Fonts | none |
-| Unbounded | `unbounded-*` | Google Fonts | none |
-| Playfair Display | `playfair-display-{latin,cyrillic}`, `playfair-display-italic-{latin,cyrillic}` | google/fonts `ofl/playfairdisplay/PlayfairDisplay[wght].ttf` and `PlayfairDisplay-Italic[wght].ttf`, **unmodified**: only the container changed (TTF to WOFF2, fontTools; every table except `head`'s WOFF2 flag and checksum is byte-identical) | **"Playfair Display"**. A subset or any other change would be a Modified Version that may not use the name, so the full fonts ship; the `latin` and `cyrillic` files are the same full font, picked by `unicode-range`. |
-
-If you add a font, prefer one without a Reserved Font Name, or ship its upstream files unchanged.
-`recipes/clean-os/fetch_fonts.py` downloads Inter and Great Vibes (OFL) from Google Fonts into a project at build time; they are not bundled.

package/skills/ghost-editor/docs/THIRD_PARTY/kinetic.md

@@ -1,22 +0,0 @@
-# Third-party sources: kinetic type (`textfx` beats)
-
-File: `scripts/lib/kinetic.mjs`. No upstream code file was copied verbatim: the blocks were written fresh for the skill's
-single-timeline GSAP contract. The table says which technique, constant or formula came from where. All sources are MIT or
-Apache-2.0 (verified 2026-10-10 from each repo's LICENSE file and `gh api repos/<r> --jq .license.spdx_id`).
-
-| repo | licence | what was taken (ideas, numbers, formulas) | used in |
-|---|---|---|---|
-| [heygen-com/hyperframes-launches](https://github.com/heygen-com/hyperframes-launches) (`HF-heygen-stripe/compositions/rotary.html`) | Apache-2.0 (NOTICE: Copyright HeyGen) | rotisserie drum geometry: words at `rotateX(-k*step) translateZ(R)`, axis pushed back by R, top/bottom mask gradient, width hugging the widest word, `rotationX` step ending on the spoken onset, no CSS filter on the 3D group | `drum` |
-| heygen-com/hyperframes-launches (`variables-launch/compositions/scene-01.html`) | Apache-2.0 | word dial with visible waiting neighbours; staggered "waterfall" entry | `drum`, `spring` |
-| heygen-com/hyperframes-launches (`claude-design-send-hyperframes-launch/compositions/s4-thesis.html`) | Apache-2.0 | character scramble: width-bounded glyph pool, `hash(step*31 + w*17 + i*7)` deterministic glyph choice, left-to-right resolve per word; `hash(i) = frac(sin(i*127.1)*43758.5453)` | `scramble` |
-| heygen-com/hyperframes-launches (`heygen-apple-motion/02-bouncy-ui/index.html`, README) | Apache-2.0 | damped spring residual `A*e^(-2*pi*zeta*f*t)*cos(2*pi*f*sqrt(1-zeta^2)*t)` with zeta = 1/3 on separate position / shape / rotation channels; volume-conserving squash `sx = e^(q/2), sy = e^(-q/2)`; 2D flip fallback `scaleY = abs(cos theta)`; one ease family (`1 - 2^(-10p)` and its mirror) | `spring`, `flipstack`, `hook`, runtime `KT.spring/EO/EI` |
-| [whaleyxbt/claude-motion](https://github.com/whaleyxbt/claude-motion) (`src/lib.ts`, `src/components/primitives.tsx`, `.claude/skills/motion-design`) | MIT, (c) 2026 whaleyxbt | mask primitive numbers (`translateY 118 %`, `rotate 5deg`, origin `0 100%`, padding/negative-margin trick); bezier curves `(0.16,1,0.3,1)`, `(0.6,0,0.9,0.35)`, `(0.65,0,0.35,1)`; line boil via `feTurbulence` + `feDisplacementMap` reseeded at 12 fps; reading-hold rule | `headline`, every mask line, runtime `KT.OUT/IN/INOUT`, `readingHold()` |
-| [bangtutorial/bang-motion](https://github.com/bangtutorial/bang-motion) (`references/techniques.md`, `references/opener-konsep.md`) | MIT, (c) 2026 Bang Tutorial | `mark()` timing (mark draws 0.4-0.6 s after the words), the under / marker / box / strike emphasis forms, the giant-cropped-word pull-back opener, the dark text "pocket" under type over busy ground | `headline`, `hook`, `pocket()` |
-| [iart-ai/tiktok-video-skills](https://github.com/iart-ai/tiktok-video-skills), [iart-ai/kinetic-typography-skills](https://github.com/iart-ai/kinetic-typography-skills) (packs of [iart-ai/motion-skills](https://github.com/iart-ai/motion-skills)) | MIT, (c) 2026 iart.ai | hook text on screen by frame 1 (no fade-up), footage punch-in as a spring, char/word stagger ranges | `hook`, `spring` |
-| [latent-spaces/brag](https://github.com/latent-spaces/brag) (`skills/brag-slim/SKILL.md`, `references/step-2-plan.md`) | MIT, (c) 2026 Shunit Haviv Hakimi | reading floor ~0.3 s per word; no crossfade between busy layouts (stagger instead) | `readingHold()`, swap mode gap |
-
-Fonts: the bundled OFL fonts in `templates/fonts` (Inter, Montserrat, Unbounded, JetBrains Mono). GSAP 3.14.2 is loaded the
-same way build.mjs always has (GSAP standard "no charge" licence).
-
-Not used: Remotion (claude-motion's runtime; company licence above a size threshold), three.js backgrounds from
-bang-motion, the HyperFrames registry text blocks (Latin-only alphabets; Cyrillic-aware versions were written instead).

package/skills/ghost-editor/docs/THIRD_PARTY/looks.md

@@ -1,22 +0,0 @@
-# Third-party code and data: transitions and film looks
-
-Everything here is MIT, except two gl-transitions ports inside the vendored xfade-easing file that are BSD-2 and BSD-3 (notices in `scripts/vendor/xfade-easing/NOTICE-gl-transitions.txt`). Nothing GPL/AGPL, nothing non-commercial, no model weights.
-
-| What | Repo | Licence | Version | What we took |
-|---|---|---|---|---|
-| xfade-easing | https://github.com/scriptituk/xfade-easing | MIT, Copyright (c) 2025 Raymond Luckhurst | v3.6.6 (commit 206e099) | `scripts/vendor/xfade-easing/eased-transitions-yuv420p-script.txt` and `xfade-easings-script.txt`, copied verbatim with its `LICENSE`. `scripts/transitions.py` reads them at run time: the standard easing expressions (Robert Penner's equations as adapted by Pohoreski and Luckhurst) and every expression-capable transition. No code from `xfade-easing.sh` or the C patch. |
-| gl-transitions (through xfade-easing's ports) | https://github.com/gl-transitions/gl-transitions | per shader; the four in the curated set are MIT (crosswarp by Eke Peter, directionalwarp by pschroen, DirectionalScaled by Thibaut Foussard, static_wipe by Ben Lucas); others reachable via `-t gl_*` are MIT, one BSD-2 and one BSD-3 | 1.71.0 | Through xfade-easing's ports in `eased-transitions-yuv420p-script.txt` (64 `GL_*` ports, shipped as upstream ships them); the BSD notices of GL_INVERTEDPAGECURL and GL_STEREOVIEWER are in `NOTICE-gl-transitions.txt`. |
-| Spectral Film LUT | https://github.com/JanLohse/spectral_film_lut | MIT, Copyright (c) 2026 Jan Lohse | 0.18.12 (PyPI) | Installed in the optional venv `~/.venvs/ghost-looks` by `scripts/setup_looks.sh`, not vendored. `scripts/gen_luts.py` calls its `create_lut()` and `load_filmstocks()` to generate `library/luts/*.cube` (data, not in git) from the film datasheets digitised in the package; `scripts/film_look.py` calls its `generate_grain()` for the grain plate when the package is importable (else numpy noise). The negative -> grain -> print chain in ffmpeg follows the tool's README node tree in our own code. |
-
-Own code (no third-party source): `transitions.py` (CSS `cubic-bezier()`/`ease*` expressions with a bisection solver, the
-easing cache, the `zoomblur`, `whip` and `softwipe` transitions, the native-xfade timestamp-warp easing, the parallel chunk
-renderer, timing, assembly, audio and the reel.json integration), `film_look.py` (halation, grain plumbing, split screen),
-`gen_luts.py` (look presets, headless driver).
-
-Python packages pulled in by `spectral_film_lut` in that venv: numpy, scipy, numba, colour-science (BSD-3), opencv-python
-(Apache-2.0), PyQt6 (GPL-3 / commercial: imported by the package, run locally only, never shipped or linked into anything
-distributed), matplotlib, imageio, networkx.
-
-Upstream bug found in 0.18.12: with print stock "Inversion (Optimized)", `FilmSpectral.cid_to_xyz` calls `input_transform()`
-with the default `upsampling_method="barycentric_nnls_all"`, which no longer exists; `gen_luts.py` patches the default to
-`"SFL upsampling"` at run time.

package/skills/ghost-editor/docs/THIRD_PARTY/matte.md

@@ -1,20 +0,0 @@
-# Third-party code, weights and ideas: person matte (`scripts/matte.py`) and cutout effects (`scripts/cutout_fx.py`)
-
-Nothing below is vendored in the repo. Weights and model code download at setup or first use into `~/.cache/ghost-matte/`
-and `~/.rembg/models/`; the venv is `~/.venvs/ghost-matte` (`scripts/setup_matte.sh`).
-
-| What | Source | Licence | How it is used |
-|---|---|---|---|
-| rembg 2.0.85 | https://github.com/danielgatis/rembg | MIT | pip dependency. `rembg.new_session(<name>)` downloads the BiRefNet ONNX files and builds the ONNX Runtime session for `--backend ort`. Its `predict()` is not used (it min/max-stretches every frame, so the alpha pumps). rembg's default model `bria-rmbg` is refused in code: BRIA weights need a paid licence. |
-| BiRefNet model code (`birefnet.py`, `BiRefNet_config.py`) | https://huggingface.co/ZhengPeng7/BiRefNet-portrait (repo https://github.com/ZhengPeng7/BiRefNet) | MIT | Loaded at run time through `transformers` `trust_remote_code` from the local cache for `--backend torch` (Apple MPS). Not modified, not vendored. |
-| BiRefNet-portrait weights | HF ZhengPeng7/BiRefNet-portrait; rembg release v0.0.0 | MIT | Default model; commercial use allowed. |
-| BiRefNet-HRSOD / BiRefNet_lite weights | HF ZhengPeng7/BiRefNet-HRSOD, ZhengPeng7/BiRefNet_lite; rembg release v0.0.0 | MIT | Optional (`--model birefnet-hrsod`, `birefnet-general-lite`). |
-| BiRefNet family weights for `cutout_fx.py`: BiRefNet_lite (default), BiRefNet (general), BiRefNet-DIS5K, BiRefNet-COD, BiRefNet-HRSOD, BiRefNet_HR, BiRefNet-matting, BiRefNet_HR-matting, BiRefNet_lite-matting, BiRefNet_lite-2K, BiRefNet_dynamic, BiRefNet_dynamic-matting | huggingface.co/ZhengPeng7/* (model cards: `license: mit`) | MIT | `setup_matte.sh` caches BiRefNet_lite; the others download on first use into `~/.cache/ghost-matte/hf` (never the repo). Model code loaded with `trust_remote_code` from that folder, not modified, not vendored. |
-| Montserrat (woff2 subsets) | the skill's `templates/fonts` | OFL-1.1 | display type in cutout_fx.py (sandwich words, cover headlines, labels) |
-| BEN2 base ONNX | https://huggingface.co/PramaLLC/BEN2 | MIT | Optional (`--model ben2-base`, ORT only). |
-| Guided filter / fast guided filter | He, Sun, Tang (ECCV 2010, TPAMI 2013); He & Sun, arXiv:1505.00996 | paper (algorithm) | Own numpy/OpenCV implementation (`fast_guided_upsample`), imported by cutout_fx.py. |
-| Blur-fusion foreground estimation | Forte & Pitie, ICIP 2021, arXiv:2006.14970 | paper (algorithm) | Own implementation of the published equations (`blur_fusion_fields`). |
-| PyTorch, torchvision, transformers, timm, kornia, einops, onnxruntime, opencv-python-headless, numpy, Pillow | PyPI | BSD-3 / BSD-3 / Apache-2.0 / Apache-2.0 / Apache-2.0 / MIT / MIT / Apache-2.0 / BSD-3 / MIT-CMU (HPND) | pip dependencies in the venv (`scripts/matte-requirements.txt`). |
-| cutout effects (light wrap, lens blur in linear light, die-cut sticker stroke, neon trace, SDF frames, sharpest-frame pick) | own code | the skill's licence | `scripts/cutout_fx.py`; textbook compositing techniques, nothing copied |
-
-Not used: BRIA RMBG 1.4/2.0 (non-commercial / paid), RobustVideoMatting (GPL-3.0), MODNet weights (CC BY-NC-SA).

package/skills/ghost-editor/docs/THIRD_PARTY/models.md

@@ -1,11 +0,0 @@
-# Models
-
-| model | used by | licence | how it gets here |
-|---|---|---|---|
-| MediaPipe Face Landmarker `face_landmarker.task` (BlazeFace short range + Face Mesh V2 + Blendshape V2, float16 v1) | `scripts/facescan.py`, `scripts/headangle.py` (gaze, blinks, head angle) | Apache-2.0 (Google; model cards) | **bundled** unmodified in `library/models/`, licence in `library/models/LICENSE-mediapipe-face-landmarker.txt` |
-| YuNet `face_detection_yunet_2023mar.onnx` | `face_track.py`, `broll_qa.py`, `headpose.py`, `headmotion.py`, the Seoul list prep | MIT (opencv_zoo) | downloaded on first use into `library/models/` |
-| SFace `face_recognition_sface_2021dec.onnx` | `broll_qa.py` (who is on screen), `own_footage.py` (only you in your own footage) | Apache-2.0 (opencv_zoo) | downloaded on first use |
-| BiRefNet weights | `matte.py`, `cutout_fx.py` | MIT (ZhengPeng7) | optional venv, see [matte.md](matte.md) |
-| silero-vad, ClearerVoice / MossFormer2 | audio QA | MIT / Apache-2.0 | optional venv, see [audioqa.md](audioqa.md) |
-| DeepFilterNet3 | `scripts/denoise.sh` | MIT / Apache-2.0 | pip into `.venv-df` on first use |
-| Whisper | `transcribe.py` | MIT (openai-whisper) | you install it (`pip install openai-whisper`) |

package/skills/ghost-editor/docs/THIRD_PARTY/runtime.md

@@ -1,13 +0,0 @@
-# Runtime tools and services (not bundled)
-
-| what | used for | licence / terms |
-|---|---|---|
-| HyperFrames CLI (`npx hyperframes`) | lint, snapshot, render | its npm licence; pin a version (`npx -y hyperframes@<version>`) |
-| GSAP 3 (loaded from jsDelivr by the composition at render time) | every timeline | GSAP standard "no charge" licence (free, commercial use included; not an OSI licence) |
-| ffmpeg, Node, Python, numpy, OpenCV, Pillow | everything | their own licences; you install them (`doctor.sh`) |
-| openai-whisper | transcription | MIT |
-| Google Gemini API (`GEMINI_API_KEY`, free tier) | B-roll people rules + line match (`broll_vision.py`, `broll_qa.py`), own-footage descriptions, reference study, AI stills, rough-cut audits | Google's API terms; frames you check are sent to Google |
-| ElevenLabs API (`ELEVENLABS_API_KEY`) | the vector explainer's narrator only | ElevenLabs terms; paid plans for commercial use |
-| macOS Vision (`scripts/ocr_text.js` via `osascript`) | the caption size/contrast check in `broll_qa.py` | part of macOS; this check needs a Mac |
-| simple-icons (fetched by slug at build time) | app logos in stan style / Seoul list icons | CC0 icons; the brands are their owners' trademarks |
-| Google Fonts (`recipes/clean-os/fetch_fonts.py`) | Inter, Great Vibes in a Clean OS project | SIL OFL |

package/skills/ghost-editor/docs/THIRD_PARTY/shaders.md

@@ -1,15 +0,0 @@
-# Third-party code and assets: shader looks and animated emoji
-
-| What | Repo | Licence | What we took / how |
-|---|---|---|---|
-| Paper Shaders core | https://github.com/paper-design/shaders (npm `@paper-design/shaders` 0.0.81) | Apache-2.0 | The `ShaderMount` class and all fragment shaders, bundled unmodified into `templates/vendor/hf-paper.js` by esbuild (`tools/paper/build-vendor.mjs`, rebuilds byte-identical). Paper's LICENSE and NOTICE ship next to it: `templates/vendor/LICENSE-paper-shaders.txt`, `NOTICE-paper-shaders.txt`. |
-| Paper Shaders React | same repo (npm `@paper-design/shaders-react` 0.0.81) | Apache-2.0 | Bundled into `hf-paper.js` with React replaced by `tools/paper/src/react-shim.js`: each component function is called once to get Paper's own params-to-uniforms mapping and presets; React is never rendered. `setMinImageSize` and `processUniforms` are ported (lightly adapted) into `tools/paper/src/hf-paper.js`. Preset values: `templates/vendor/paper-shaders.json`. |
-| Fluent Emoji Animated | https://github.com/microsoft/fluentui-emoji-animated | MIT (c) Microsoft | 48 APNGs (256 px) downloaded by `scripts/emoji_fetch.py` (sha256 checked against the Git LFS pointer) and re-encoded to VP9-alpha WebM in `library/emoji/` (not in git). |
-| Fluent Emoji (static 3D) | https://github.com/microsoft/fluentui-emoji | MIT (c) Microsoft | 16 static 3D PNGs for symbols with no animated version (git blob sha1 checked). |
-| esbuild | npm `esbuild` | MIT | Dev dependency of `tools/paper` only (bundling). |
-
-Our own code: `scripts/lib/shader-layer.mjs` (looks, layers, the word scene), `scripts/lib/emoji-pop.mjs`, the HyperFrames seek
-driver in `tools/paper/src/hf-paper.js` (redraws from the `hf-seek` time, never requestAnimationFrame), `scripts/framecheck.py`.
-
-The emoji fetcher writes Microsoft's MIT licence files next to the media (`library/emoji/LICENSE-*.txt`); keep them with the
-files if you redistribute them.

package/skills/git-story-film/SKILL.md

@@ -1,220 +0,0 @@
----
-name: git-story-film
-description: "Direct an animated film of about two minutes from a software repository's Git history. Select real commits, releases and reverts; develop the story, character and visual style; build an outline, cast sheet, storyboard and HTML review studio; then render a 4K/60fps MP4 after approval. Use for repository or product promos, launches, anniversaries, release stories, codebase timelines, character-led project histories, and requests such as \"make a video about our project\" or \"show how this product grew.\" Not for UI screen recordings, slide decks, or commit-statistics charts. Runs in AICUT: the film is film.html, a page per shot, drawn by `aicut look` and `aicut render`."
-license: Apache-2.0
-source: "EverMind-AI/Raven"
----
-
-# Git story film
-
-## Running it in AICUT
-
-AICUT's film is `film.html`: it cuts pages, footage, pictures and sound together on tracks. A page is a web page that
-sets `window.film = { frame(t) }` and draws that moment — no framework, no build step, no composition object, no
-`registerRoot`. Five shots are five pages cut together, not one page that changes scene, so each shot stays trimmable
-and replaceable in Studio.
-
-```sh
-npx -y @autoaiai/aicut@latest          # the manual: the format and the commands, always current
-npx -y @autoaiai/aicut@latest open .   # make this folder a film and print Studio's address
-npx -y @autoaiai/aicut@latest look     # a contact sheet of the whole film, its sound, and a determinism check
-npx -y @autoaiai/aicut@latest look 4.5 # one frame, full size — look at the PNG, not at the code
-npx -y @autoaiai/aicut@latest render   # the MP4, beside film.html
-```
-
-Run the manual before the first film: it is authoritative for the format, and this skill is about the method above it.
-
-Where the words below name another program's part, it becomes this:
-
-| elsewhere | in AICUT |
-|---|---|
-| composition, timeline, project file | `film.html`; tracks and clips inside it |
-| a scene, a sequence, a slide | a page (`pages/<name>.html`) cut onto a track |
-| `useCurrentFrame()`, `interpolate()` | `frame(t)` — the picture depends only on `t`, never on state or `Date.now()` |
-| a still, a screengrab | `aicut look <t>` |
-| `render` / `build` / `export` | `aicut render` (the person can also export from Studio) |
-| a dev server, a preview window | `aicut open` and the Studio address it prints |
-
-Hard rules that hold whatever the method says: the picture depends only on `t`; every value in the person's prompt is
-binding; look at the frames `look` writes before saying a film is done.
-
-
-The user gives you a GitHub link. You direct a film of about two minutes on
-how that product grew up. The git history is the footage: it supplies the
-order, the dates and the proof. Your job is to turn it into a story a newcomer
-enjoys and learns from: what the product is, how it got here, what it can do
-for them.
-
-Directing means you make the calls and show them, and the user approves.
-You own three things:
-
-- **Data.** Which facts carry the film. Out of thousands of commits, pick
-  the first commit, the release, the busiest day, the reverts and the hero
-  feature, and put real hashes, dates and counts on screen. Everything shown
-  is true and sourced; illustrative numbers are marked.
-- **Story.** The protagonist, whose voice the captions are, the form the
-  growth takes on screen, and the arc: setup (the first commit, the problem) → development
-  (abilities arrive) → turning point (the revert, the rewrite, the busiest day) → resolution (where
-  it landed, and what it means for the viewer).
-- **Look.** The art style and how it changes: a look per act with a story
-  reason, the palette, the protagonist's design, type and sound.
-
-A mascot hatching and walking along the timeline is one form, not the only
-one. The same growth can be staged as something growing in place, a machine
-being built, game levels, a user's days, or a form the product suggests
-(`references/forms.md`). Choose for each product.
-
-The engine in `engine/` (MIT, adapted from alesha-pro/tools hand-drawn-canvas-animation,
-https://github.com/alesha-pro/tools/tree/main/skills/hand-drawn-canvas-animation; keep its LICENSE with
-it) draws and animates, and `engine/studio-ui.js` turns
-any film page into the review studio. `examples/raven-story/` is a finished
-journey film (128 s, made for EverMind's Raven, with Chinese files in `zh/`).
-Reuse their code; do not start from a blank page.
-
-The instructions and default examples use English. Keep dedicated Chinese film
-assets in `zh/` and Chinese UI labels as localization data; translating those
-into English would remove the Chinese cut. Choose the output language for the
-audience, and use `--ui zh` only for a Chinese storyboard interface.
-
-## Setup
-
-Clone the repo outside the film folder (`git clone --filter=blob:none <url>`)
-and research it before step 1: run `python3 scripts/git_timeline.py <repo>`,
-read the README feature list and docs, and note the first commit, public
-release, busiest day, reverts, hero feature, built-in agents/modules and real
-integrations (`references/story.md` §1, §6). Ask about audience, languages and
-where the film will be shown if you don't know.
-
-Set up the film folder, also outside the product repo
-(`references/production.md` §1):
-- copy `engine/*`, `scripts/render-hq.mjs` and `scripts/frames.mjs`;
-- run `npm i` and `sh scripts/fetch_fonts.sh <dir>`;
-- run `git init` with `node_modules/` and `out/` ignored.
-
-For a journey, start from the example. For a Chinese film take
-`raven-cast.js`, `raven-world.js`, `raven-route.js` and `raven-flight.js` from
-`examples/raven-story/zh/`, and the rest (`raven.html`, `raven-score.js`,
-`cast-sheet.html`) from `examples/raven-story/`; for English take everything
-from the top level. For another form, keep the example's film HTML, writer and
-score as the skeleton and write the world code fresh. Every film HTML loads
-`studio-ui.js` right after `core.js`. The bundled example links to
-`../../engine/`; when copying it into a flat film folder with the engine,
-change those script paths to the local filenames.
-
-Captions carry the voice and the score is synthesized; there is no voiceover.
-If the user asks for voiceover, say so and offer to leave timed caption
-gaps for a recorded track.
-
-## The director's workflow
-
-Seven steps, in this order. Each ends with something the user can look at, and
-you stop for their approval before the next. Story, character and pacing are
-cheap to change on paper and expensive after a 4K render. Templates, shot
-fields and checks are in `references/directing.md`.
-
-1. **Outline: story direction.** Tell the user what story you'll tell and what it
-   highlights. Start `OUTLINE.md` with a logline, the form and why it fits, the
-   protagonist and caption voice, the four-part story arc in four lines, the 3–5
-   things the film will make the viewer remember (and which gets the most
-   time), the numbers it stands on, and the look in a line or two.
-2. **Character: expressions and actions.** Design the protagonist (and companions) in the
-   chosen look. Render `cast-sheet.html` with 6–12 expressions, key poses for
-   each action the story needs (walk or hop cycle, jump, fly, talk, point,
-   carry), every material they appear in, and a 1×/2×/3× size check
-   (`references/craft.md` §1–2, §8).
-3. **Outline: the full film.** Expand `OUTLINE.md` to the whole film, about 2:00: acts
-   and looks with reasons, recurring devices, one line per beat (time, what
-   happens, feature, commit), the ending, and claims to double-check.
-4. **Storyboard: the first 30 seconds.** Write the shots for the first 30 seconds into
-   `shots.json`, draw one keyframe per shot with the real cast and world code,
-   and open the storyboard page:
-   `python3 scripts/storyboard.py shots.json --frames storyboard/frames --out storyboard/storyboard.html --ui en`.
-   These 30 seconds fix the tone, pacing, voice and look for the rest.
-5. **Build the full film.** Write the rest of the shots and animate the whole film.
-   When a shot needs a new expression or action, add it to the cast sheet and
-   show the user; they may ask for more at any time. Check your own work with
-   the preview tools (`references/production.md` §2). Don't render.
-6. **Studio preview.** Write the shot data and open the film page, which
-   `studio-ui.js` turns into the studio:
-   `python3 scripts/storyboard.py shots.json --js shots.js --out storyboard/storyboard.html`, then
-   open the film HTML. The film is on top with a scrub track (acts as bands,
-   shots as cells, drag anywhere, play from any point with the score in sync)
-   and the storyboard below, one frame per shot drawn live from the film;
-   click a frame to jump there. Iterate on the user's notes until they
-   approve. Still no MP4.
-7. **Final delivery.** Only after approval in the studio, deliver with
-   `node render-hq.mjs film.html --width 3840 --fps 60`, one language at a
-   time (a 1080p24 file first, `--width 1920 --fps 24 --no-cover`, only if the
-   user wants a quick file to share). By default the closing frame is also
-   written as frame 0, so players and feeds show it as the cover, and it
-   flashes for one frame before the film starts. Keep it for social uploads;
-   pass `--no-cover` for anything played inline or looped. Verify with ffprobe
-   and a full decode, and report duration, size, resolution and what you
-   actually looked at. Then swap the storyboard's keyframes for real frames
-   (`node frames.mjs film.html shots <seconds>`).
-
-Other languages live on a branch. Translate as native copy, not literally;
-localize references (WeChat → Slack, only if the product has Slack); and
-re-check every text-heavy frame in the studio after translating, because
-longer lines collide.
-
-## Things that matter more than they look
-
-- **No render before the studio.** The user approves the film in the studio;
-  the MP4 comes after. A 4K render is the most expensive place to find a
-  problem.
-- **Story before spectacle.** Each act teaches the viewer one thing, the turning point is the
-  peak, and the ending pays off something set up in the first 20 seconds.
-- **The form's grammar holds for the whole film.** A journey never cuts and
-  only moves right; growth stays in place and pulls back; levels cut only
-  between levels. Break the grammar once and the film turns into a slideshow
-  (`references/forms.md`).
-- **Literal mapping.** One token per commit, one milestone per release, one
-  setback per revert, with real daily counts. That is what makes it this
-  repo's story.
-- **Two clocks.** Position every output frame; poses on ones/twos. Mixing
-  them up makes the protagonist stutter at 60 fps (`references/craft.md` §3).
-- **Fonts must be licensed for video.** Never ship a system font into a
-  published film; bundle OFL faces (`references/text-and-fonts.md`).
-- **Truthful beats.** Real hashes and dates; mark illustrative numbers; don't
-  let a development tool read as runtime magic; only list integrations that
-  exist.
-- **Native copy in every language**: spaces between Chinese and Latin text,
-  official product casing, and a caption that says something new each time.
-- **One render at a time.** Parallel 4K renders crash Chrome workers.
-
-## Reading map
-
-| When | Read |
-|---|---|
-| Each step's document, shot fields, the studio, checks before a hand-off | `references/directing.md` |
-| Choosing how the growth is staged (journey, growth, building, levels, companion...) | `references/forms.md` |
-| Evidence, the setup, development, turning point and resolution arc, beats, devices, looks per act, fact checks | `references/story.md` |
-| Drawing the protagonist, poses, motion, props, seams, companions, pixel art | `references/craft.md` |
-| Fonts, captions, translation, spacing, text layout checks | `references/text-and-fonts.md` |
-| Folder layout, preview commands, 4K delivery, sound, pitfalls | `references/production.md` |
-| How the pieces fit in real code (a journey) | `examples/raven-story/` (start at `raven.html`) |
-| A text-only storyboard reference with bilingual captions; view live frames in the film studio | `examples/raven-story/storyboard/storyboard.html` |
-| Engine API (palettes, marks, cels, exposure, camera) | comments in `engine/core.js`, `engine/cels.js`, `engine/studio.js` |
-
-## Scripts
-
-| Script | Does |
-|---|---|
-| `scripts/git_timeline.py <repo>` | daily counts, totals, busiest days, feat/perf commits with PRs, reverts, tags → `timeline.json` |
-| `scripts/fetch_fonts.sh <dir> [zh] [en]` | downloads OFL fonts with licences (zh: wenkai/xiaolai/kuaile; en: comicneue/patrickhand/shortstack/architectsdaughter) |
-| `scripts/frames.mjs film.html load|shots|score` | page errors and frame count; stills at given seconds; score rendered with per-second levels |
-| `scripts/storyboard.py shots.json [--frames dir] [--embed] [--js shots.js] [--ui zh] [--length s]` | the storyboard page: a storyline strip (acts, durations, time ticks), then the shots grouped by act, text-only or with stills (`--embed` makes one sendable file). `--js` also writes the shot data the studio reads |
-| `engine/studio-ui.js` | loaded after `core.js`, makes the film page the studio: scrub track with acts and shots, playback from any point with the score, keys, live storyboard frames; renders (`?bare`) never build it |
-| `scripts/render-hq.mjs film.html [--width 1920/3840] [--fps 24/60] [--no-cover] [--param k=v]` | streaming parallel render straight into ffmpeg, score muxed, closing frame as cover unless `--no-cover`; writes `out/<name>-<height>p<fps>.mp4` |
-| `engine/render.mjs` | preview modes at 12/24 fps only: `--grid N`, `--only i,j`, `--strip start,count` (frame indices) |
-
-Requirements: Node 22+, Google Chrome, ffmpeg (Homebrew path may need
-exporting), Python 3.8+.
-
-## Source
-
-Rewritten for AICUT from [`EverMind-AI/Raven`](https://github.com/EverMind-AI/Raven)
-(Apache-2.0), skills/git-story-film/SKILL.md. The method is that skill's own; the way it runs is AICUT's.
-
-The repository holds 7 skills; this is the one its entry stands for.

package/skills/git-story-film/examples/raven-story/FONTS.md

@@ -1,8 +0,0 @@
-# Fonts used by the example
-
-All SIL OFL 1.1, fetched with `sh scripts/fetch_fonts.sh <dir> wenkai comicneue`.
-
-| File | Family | Used by |
-|---|---|---|
-| `LXGWWenKai-Medium.ttf` | LXGW WenKai v1.522 | Chinese cut (all text) and CJK fallback in the English cut |
-| `ComicNeue-Regular.ttf`, `ComicNeue-Bold.ttf` | Comic Neue | English cut |