Build guide · the video skill

Give your agents the power to make video.

In every fleet I've built, video is the bottleneck. An agent can write you a killer script in ten seconds — and then it stops, because turning that script into a finished short has always needed a human, a timeline editor, and an afternoon. This guide kills that handoff. You wrap the whole pipeline in a single skill, and a one-line brief comes back as a rendered mp4 — voiceover, synced captions, music, the works.

The unlock is one idea: video as code an agent can author. I render every one of my shorts this way. Here's the recipe so your agents can too.

When video is something any agent can output instead of something only a specialist can produce, it stops being a bottleneck and becomes a default deliverable.

◇ Jump to the copy-paste recipe See the pipeline →

Part 01 · the engine

Why Remotion: video an agent can write.

Most video tools are timelines you drag clips around in. You can't hand a drag-and-drop timeline to an agent. Remotion is different — it's video composed in React/TSX. A composition is real code: components, props, data-driven scenes, deterministic frames. You define a 1080×1920 canvas, a duration in frames, and a tree of scenes — a title card, a lower-third, kinetic word-by-word captions, a brand bumper — and Remotion renders it headless to an mp4.

That's the whole reason this works. Because the video is code, an agent can write it: scenes, timing, captions, brand colors, all as props it fills in. It renders without a human ever opening an editor. Everything below — the voiceover, the captions, the music — is just generating the assets that feed the composition, then calling render.

Part 02 · the pipeline

Brief in, mp4 out — six steps.

The skill is a straight line. Each step produces an artifact the next one consumes. Nothing here is magic; it's six tools wired in order.

brief → script → vo + captions + music → remotion render → mp4 → store

The agent only writes the script and picks a recipe. Everything downstream — VO, captions, music, render, storage — runs as wired tools. The dashed line is the one cross-link that matters: the voiceover audio is what Whisper times the captions against.

Part 03 · the voice

Two voices: free for volume, paid for showcase.

Don't pick one TTS — wire two tiers and let the brief choose. Most drafts never need to cost a cent; the pieces that go out the door get the expensive voice.

Tier A · default

Free, local, on-device

e.g. Kokoro — runs on your own hardware

The workhorse. Renders fast, costs nothing, never hits a rate limit. Use it for every draft, every internal cut, and any high-volume lane where you're shipping dozens of variations a day.

$0 per render, no API quota
Good enough for drafts and b-roll narration
Lets agents iterate without a budget conversation

Tier B · showcase

Paid, expressive, cloud

e.g. ElevenLabs — hosted premium voices

The closer. Reach for it only when the piece actually goes to a client, a launch, or the top of a funnel — anywhere the voice is the product. The expressiveness is worth the cents per render when it's the face you're showing.

Natural, emotive delivery that reads as human
Per-character cost — spend it on finals, not drafts
Pin one voice ID per brand for consistency

The skill just exposes a voice_tier param. Free by default; the agent escalates to paid only when the brief says "showcase." Same pipeline, one switch.

Part 04 · the sync

Captions that actually sync.

On-screen captions are what make short-form watchable on mute. The trick isn't generating the text — you already have it, it's the script. The trick is timing.

Run the rendered voiceover audio through Whisper and ask for word-level timestamps. Now every word has a start and end frame, so the caption highlights in perfect lockstep with the VO — the kinetic, word-by-word look that holds attention. Remotion just reads those timestamps and lights up each word on its frame.

One catch worth teaching: speech models mishear homophones ("their" vs "there," a brand name, a number). You already know the true text — it's the script you wrote in step two. So align Whisper's output back against the known script: keep Whisper's timings, but correct its words to your script. You get machine-perfect timing and zero typos. That alignment step is the difference between captions that look automated and captions that look authored.

Part 05 · the multiplier

Recipes are the force-multiplier.

Authoring a bespoke composition every time is slow and inconsistent. The real leverage is a recipe library — parameterized templates the agent fills in instead of building from scratch.

What a recipe fixes· structure

One recipe = aspect ratio + duration band + scene grammar + shared components

Each recipe pins the decisions that should be consistent and exposes the few that should vary. A client-promo recipe is locked to 16:9, 45–65s, a fixed scene order (hook → three beats → CTA bumper), and a set of shared Remotion components. The agent supplies the params — client name, the three beats, brand color, logo — and the output is on-brand and fast instead of bespoke and slow.

Starter library· lift these

client-promo · explainer · tribute · readme-walkthrough · b-roll

These are the five I lean on. client-promo (landscape, punchy). explainer (vertical, captions-forward). tribute/testimonial (longer, slower, music-led). readme-walkthrough (screenshot-to-animation for a repo or product). b-roll (short, silent, no VO). Each one fixes its own aspect ratio, duration band, and render defaults so the agent never re-decides them.

Keep it DRY· shared parts

Share components across recipes — lower-thirds, kinetic captions, brand bumpers

Build the reusable pieces once and import them into every recipe: the lower-third, the word-by-word caption component, the intro/outro brand bumper, the render presets. Fix a bug in the caption component and every recipe inherits the fix. This is the whole advantage of video-as-code — you refactor your video library like you refactor a codebase.

Don't re-render· catalog

Catalog every rendered asset so agents reuse instead of regenerate

Log every finished mp4 — its path in storage, the recipe and params that made it, the project it belongs to. Next time an agent needs that promo, it finds the existing render instead of burning compute and a paid VO making it again. Render is the expensive step; a catalog turns it into a cache.

Wire it once. Then video is just another thing your fleet outputs.

The point isn't one slick video — it's that any agent can call this skill. Video stops being a specialist bottleneck and becomes an on-demand deliverable, the same way an agent already writes copy or ships code. Build the pipeline, seed a few recipes, catalog the renders. Then a one-line brief comes back as a finished short — at the volume only an agent can hit.

Hit "Copy the recipe" up top to lift the whole skill spec — the pipeline, the stack, and the recipe pattern — straight into your own agent.