Communitygithub.com

create-short-video

Use when making a short vertical (9:16) Thai explainer clip for Reels, Shorts or TikTok — news, tool review or demo, concept, how-to. Hand-drawn style, voice recorded by the owner, rendered with Remotion.

What is create-short-video?

create-short-video is a Claude Code agent skill that use when making a short vertical (9:16) Thai explainer clip for Reels, Shorts or TikTok — news, tool review or demo, concept, how-to. Hand-drawn style, voice recorded by the owner, rendered with Remotion.

Works with~Claude Code~Codex CLI~Cursor
npx skills add https://github.com/savepong/skills/tree/HEAD/skills/create-short-video

Ask in your favorite AI

Open a new chat with this agent skill pre-loaded.

Documentation

Explainer short (9:16, ~60-90 s, Thai voice, any topic)

Runbook for a vertical explainer: news, a tool or repo review, a concept, a how-to, or a demo of something the owner built. Build mechanics and style porting live in the create-video skill; load it before step 4. Voice engines and their quirks: references/voice-options.md.

Requires: bun, ffmpeg/ffprobe, python3, and for measure_ink.py Pillow + numpy (uv run --with pillow --with numpy ...). Thai prose polish uses the kien-thai and kode-thai skills (npx skills add chakrit/kien-thai).

Defaults (no need to re-ask)

  • 9:16, 1080x1920, 8-10 beats, about 60 s (Shorts allow up to 3 min; a self-read script usually lands at 80-90 s)
  • Style: the last Atlas style the owner approved. Kits in the template: Ballpoint (src/style/ballpoint.js, window.BP) and Whiteboard (src/style/whiteboard.js, window.WB). A new style needs the pasted Futurepedia AI Style Atlas prompt.
  • Visuals drawn in code. Real logos only for the subject and as source credits. No news photos. Exception: when the subject is a live web demo, mix real browser screenshots you capture yourself for the "see it work" beats with drawn beats for the mechanism.
  • Voice: the owner records it himself. Hand over read-script.md, he drops a recording, scripts/split_recording.py or scripts/assemble_takes.py cuts it into beats. Synthetic voices only if he asks (references/voice-options.md).
  • No emoji, no API keys, cite sources.

Pick the story shape first

Shapes and beat templates: templates/story-shapes.md.

ShapeUse forLive re-check
NewsSomething happened, may still be happeningYes, before render
Tool review / demoA repo, app or product: what it does and the catchRun it yourself
ConceptHow something worksNo
How-toSteps to do somethingNo, but test the steps

Steps and approval gates

  1. Topic pick. Prefer stories with a mechanism to draw over who-did-what news; skip scandal/gossip and fresh lawsuits with nothing yet to explain. Give 3-4 ranked picks with why, and say they come from headlines not yet verified. Check date first.
    • Freshness: a news clip that cannot ship within ~1-2 days of the event is stale; say so before investing. Offer instead: wait for the primary follow-up (e.g. a post-incident report), or recast as an evergreen Concept clip using the incident as an example.
  2. Research. Read the primary source (repo, docs, status page, official post) plus outlets. Keep three labels apart: verified by me / claimed by the source / my read.
    • Handoff issues/briefs: verify every "current state" claim against live state (is the PR merged, is the URL up, is the file where it says). Read files from GitHub (gh api repos/O/R/contents/PATH -H 'Accept: application/vnd.github.raw') rather than pulling or checking out the owner's local repo.
    • Demos: run the real thing in a browser, reproduce any stated numbers, and measure asset sizes (curl -sIL ... content-length) instead of trusting estimates hardcoded in source. Report mismatches in a claim/actual table.
    • "No server" / privacy claims: list the page's network requests (performance.getEntriesByType('resource')). Hosting analytics beacons (e.g. Cloudflare /cdn-cgi/rum) show up in a DevTools shot; filter to fetch or explain them on screen.
  3. Script from the shape's template, numbers as words, length hit by cutting words never by speeding up. Write read-script.md (numbered beats, one breath each).
    • News = inverted pyramid: beat 1 is a complete lead (who, did what, when) that stands alone; then most to least important; attribute every claim ("X ระบุว่า", "ด้าน ... เชื่อว่า"); no opinion or moral in the voice, takeaways go in the caption.
    • Credits and URLs the brief requires go in the last beat and on screen.
    • Polish with kien-thai via the kode-thai loop: audit, cite rule slugs, fix, re-read, repeat until one pass finds nothing. Spoken explainers by the owner = personal-blog register, male: ผม, ครับ only at sign-off, คุณ for the viewer, เรา for the team. Recurring fixes in model drafts: เรา drifting between builder and viewer (deixis-continuity), ทิ้ง for "drop" (verb-calque), dangling "ไม่ใช่ X เป็น Y" without ต่างหาก (f4/targhak-closure), mixed เมกะ/เมกะไบต์ (one form per term), repeated ก็เลย/พอ across adjacent beats.
  4. Voice -> GATE 1 (owner listens). Recording instructions (QuickTime > File > New Audio Recording, Quality Maximum, not AirPods; or iPhone Voice Memos + AirDrop): pause ~2 s between beats, clearly longer than a breath (self-recorded gaps came out 0.6-0.9 s, the same as mid-sentence breaths, so no threshold split them into exactly N); re-read a fluffed beat immediately; read English names naturally. Offer to check a 10 s test of beat 1 first.
    • Find the recording: he may drop several files (takes) into a subfolder and say only "done". Search the project folder recursively plus find ~ -mmin -30 -iname '*.m4a', then per file report duration, volumedetect mean/max and silence map before cutting. Quote paths in loops (Voice Memos names contain spaces).
    • One file: python3 split_recording.py <file>; it stops unless it finds exactly N beats. If no threshold gives the right count, do not force a cut: ask for rough beat start times and snap each cut to the nearest silence.
    • Several files / retakes: ask which file is which beat, then assemble_takes.py with a takes.json of {file, start, end} per beat. Split a multi-beat file at its longest silences and sanity-check each cut against the speaking rate of the single-beat files (~11 Thai chars/s); flag beats whose voiced time is >20% short (possible skipped words). It level-matches every beat (beat_lufs), shortens mid-sentence pauses to max_pause by cutting the middle of each pause (ffmpeg silenceremove barely does this), then compresses before loudnorm (speech crest ~22 dB; lands near -18 LUFS at TP -1.5).
  5. Storyboard -> GATE 2 (owner approves). Stills at each beat's finish time as one contact sheet plus full-res PNGs. List anything assumed or illustrative.
    • Measure framing from pixels, never by eye (eyeballing once missed a uniform right shift). scripts/measure_ink.py --z .. --camx .. --camdy .. stills... fits each beat's ink inside the TikTok/Reels safe box x 140-940, y 230-1500 (content spilled under the app UI otherwise) and prints corrected CAMXS/CAMDY/ZS; one re-measure pass leaves >= 8 px margin. Ship a storyboard-sheet-safezone.png with the unsafe bands tinted (ffmpeg drawbox).
    • Take each still right after that beat's LAST mark (its IO end), not at next-beat-start minus a constant: a beat that finishes early is already scrolling away and measures wildly off.
    • Showing OCR / text-recognition errors: loopless Thai faces (Sriracha, Mali, most UI fonts) draw น like Latin u and บ/ป like U, so a wrong read such as Qนotation LOOKS correct on screen and to vision checks. Verify strings via JS text, not screenshots; draw wrong reads in Itim (looped, OFL, bundled in the template) and ring the flipped glyph using measureText. WB.mText takes a hex colour: pass B.hexOf(RED), not the RGB array.
  6. Render and verify. bun run render <slug>. Frames == duration x 30; voice LUFS in the mp4 == voice file (narration muxed); score ~20+ dB under the voice; mid-draw stills (not beat ends) to check the pen sits on the ink and nothing leaves the safe box. For News, re-read the primary source first and update the on-screen time stamp.
  7. Deliver to ~/Downloads/<slug>-video/: full mp4, -small.mp4 (CRF 24), score wav, captions from templates/captions.md. Give absolute paths.
  8. Publishing. Never publish publicly for the owner. If uploading for him, set Private and tick YouTube's "Altered or synthetic content" only when the voice is synthetic; he presses Publish. Otherwise give a copy-paste kit: file path, title, description, settings to tick.

Scaffold

scripts/new_short.sh <slug> [--out dir] [--from template] [--kits dir] [--no-install] copies the bundled template/ (Itim + Thai fonts, safe-box camera, render/stills scripts, reference videos) to ~/Videos/shorts/<slug>/ (or $SHORT_VIDEOS_DIR), keeps earlier videos as _reference_* to adapt, and copies the voice tools into <slug>/voice/. Dependencies install automatically (--no-install skips). Pass --from <a finished project> to start from your latest approved clip instead.

The Futurepedia Style Atlas material (H library, SOUND kit, prompts and the Ballpoint/Whiteboard kits ported from their reference code) is NOT in this skill — it belongs to Futurepedia. new_short.sh copies it from $CODE_MOTION_KITS (default ~/Videos/shorts/_kits/{src/lib,src/style,prompts}). Without it: save the Atlas prompt to proj/prompts/, bun run extract-kit prompts/<style>.md, then port the style (create-video). Never commit those files.

Related Skills