Shorts director
A short is judged in two seconds by someone who owes it nothing. Everything in this skill exists so that the first cut you hand over is one the owner would post: the idea chosen from three, the words visible on screen as well as heard, no fault in the frame that they spot before you do, and a sound mix they can play on a phone in a danfo. Quality comes first. Cost is shown before every spend, never used as a ceiling.
The skill directs; the repo's engines do the work. The engine tracks (the Remotion kit, the clip and image tools, the niche libraries, the collage camera) and the tools added since (mouth_sync, the shared cards, the caption chunker, the checks) are render paths this skill picks between, beat by beat, rather than one track per video.
Before anything: read the config and the bible
Identity lives in config, never in the skill. Read director.config.json at the repo root for the brand, the corner mark, the people a film may name, link tagging, house rules, the providers and their prices, and the mastering target. Visual identity (colours, fonts, logo) is remotion/src/brand.ts and remotion/src/fonts.ts. If the short belongs to a series, read its bible in ai-shorts/series/<series>/BIBLE.md before writing a line; if it is a new channel or cast, the series mode writes one first. Then read the last few entries of RUNS.md and LESSONS.md, because the faults of the last run are the cheapest ones to avoid in this one.
The modes
| Mode | Its job | Read |
|---|---|---|
| series | the bible for a recurring cast or a channel style | modes/series.md |
| writers-room | three divergent premises judged as the buyer, then the script | modes/writers-room.md |
| produce | voice, animatic, stills, clips, mouths, composition | modes/produce.md |
| sound | beds, cues, the mix and mastering | modes/sound.md |
| finish | the master, the phone copy, the text check, the posting pack | modes/finish.md |
| bakeoff | trying models against each other before one becomes the default | modes/bakeoff.md |
| effects | /effects <name>: preview an effect, place it on a beat, or build a new one into the library | modes/effects.md |
| improve | after every short: score the run, land the lessons, propose skill changes | modes/improve.md |
A new short runs them in order: series (only if the bible is missing), writers-room, produce, sound, finish, improve. The owner can enter at any mode by name, and a request that clearly sits in one mode (a voice re-take, a caption fix, a new hook) goes straight there.
The fast lane. When the series bible exists, the format is proven and the engines are known, skip nothing but shorten everything: one premise pass instead of a long one, and stills reused from the series where they fit. The gates and the checks stay exactly as they are; only the rounds get fewer.
The gates that hold in every run
These are where the owner's eye goes, in order, and where the work stops until they answer. Each one is cheap to change at and expensive to change after.
- The premise. Three, written and judged as the buyer, with a pick and the reason. Nothing is voiced before one is chosen.
- The voice. Lines in the owner's words where they gave them. For anything in pidgin or a local name, auditions they choose by ear (
references/voice.md). A picked take is seated exactly, never re-generated. - The animatic. Stills on the real voice timeline with the code-drawn gags, captions and red tags where clips will go. No clip is generated before the animatic is approved.
- The spend. Every generation step states its credits or dollars first (
references/stills-and-clips.md), and the total is reported after. - The cut. The owner watches the rough cut before mastering, and the finished file before anything is posted.
Between the gates, check your own work the way references/qa.md describes, so the owner's eye is spent on taste and never on a fault a scan could have caught.
Formats
A format brings its own grammar (how the beats run), its own visual world and its own default render paths. The ones proven so far each have a reference:
references/formats/skit.md: a recurring cast in dialogue (proved by Akpos and Segs, episodes 1 and 2)references/formats/documentary.md: a narrator over photoreal scenes with sourced figures (proved by Qodeup mini documentary 1)
The engine formats stay available too: code-drawn explainers through make-short and vidtsx-2d-generator, collage through make-vox. The director can route a whole short to one of them, or borrow a beat from one.
When the requested format isn't proven yet
If the owner asks for a kind of short none of the formats covers (a listicle, a before-and-after, a talking-object, a news recap), say so plainly rather than forcing it into the nearest format; a skit grammar on a news recap produces something that feels off in a way they will notice. Offer two paths. The first is to set it up properly: run it as a proving run, and once it ships, write references/formats/<name>.md from what the run proved, through os-skillify, so the next one starts from a grammar instead of a guess. The second is to make it once with the closest format as scaffolding, with the gaps named, and save nothing. If the same new format comes up a second time, take the first path; a format that keeps being improvised never gets a grammar, and the skill quietly narrows what the owner asks it for.
How it gets better
Every short ends in the improve mode. It scores the run against the outcomes in references/outcomes.md, logs it in RUNS.md, writes what was learned to LESSONS.md, and when a fault has now happened twice or a technique has proved itself, proposes the change to this skill in general terms, as a diff the owner approves. Git is the way back.
References
references/outcomes.md: what the skill must produce, with numbers, how each is checked, and the loops that improve itreferences/render-routing.md: choosing, per beat, between code, a cut-out, a generated clip and the engine tracksreferences/voice.md: voices, casting, auditions, pidgin and names, seating a picked takereferences/stills-and-clips.md: generating stills and clips, prompts that hold identity, re-rolls, pricesreferences/mouths.md: redrawing mouths on generated clips with mouth_sync, including per-shot settingsreferences/facts.md: the fact gate for anything stated as true on screenreferences/effects.md: the jobs a beat can have, the effect that serves each, and the restraint rules that keep a short from becoming a demo reelreferences/qa.md: the checks between gates, and what each one has caught beforereferences/formats/: one file per proven format
Related skills
make-short,vidtsx-2d-generator: the code-drawn explainer track and its crash rules, used as a render pathmake-vox: the collage trackmake-ai-short: the generative track, whose clip discipline this skill carries forwardsuggest-sfx: sound cue auditions, called from the sound modeos-writing(Personal OS): hooks and captions for posting, called from the finish modeos-skillify(Personal OS): writing a new format reference after its proving run