Skills / video
video
Make a short, scroll-stopping video from a plain request, rendered with Remotion on your own computer. Three formats done well - a tech explainer ("explain how rate limiting works"), a kinetic-type short (bold words that hit on the voice) and an app showcase (a recorded demo of your own app). Use whenever the user asks to make a video, short, reel or demo. Don't use to upload a video.
- Author
- agentsaisle
- Latest
- v0.29.0
- Aisle
- video
- Works with
- Claude Code
- Source
- HazemSalah29/agentsaisle
- Stars
- 2
Install
Install video
npx @agentsaisle/cli add videoInstalls into your project's skills folder. Add --global to install it for every project.
Before you start
- Install the skill
Run the install command above. Open Claude Desktop (Code tab, Local) or Claude Code in an empty folder, type /video and say what you want.
- Set up the Remotion engine once
The first time, /video writes the engine's package.json to ~/.aisle-video/remotion and shows two commands to run there: npm install, then npx remotion@4.0.540 browser ensure (about 350 MB in all). Remotion is free for individuals and companies of up to 3 people; larger companies need a Remotion company licence.
- Add the free voice
pip install kokoro-onnx soundfile, then put kokoro-v1.0.onnx and voices-v1.0.bin (from the kokoro-onnx GitHub releases) in ~/.aisle-video/kokoro. /video doctor shows what is ready. Your own recording works too, with no setup.
What it can do
Type /video and say what you want. Claude picks the tools your request needs.
- Explain an idea, algorithm or API in a 30 to 45 second short, with code, diagrams and real simulations
- Make a kinetic-type short where bold words hit as they are spoken, with slab cuts and sound design
- Make a showcase of your own app: a scripted recording with a camera that follows the action
- Write the hook and script to a short-form quality bar, then review each draft against a score card
- Voice it locally for free (Kokoro), with burned captions, an original music bed and sound effects
- Check the result, then prepare it for YouTube Shorts, Reels, TikTok or LinkedIn (nothing is uploaded)
Built from 9 parts, each audited on its own
Runs scriptsThis skill includes scripts
Your agent may run these scripts on your machine. A maintainer has read every one in full, and the CLI lists the files when you install. You can read them on GitHub first.
Read the files on GitHubAudit
Checked with aisle audit (version 0.29.0). Run it yourself: aisle audit video
- Critical
- 0
- High
- 0
- Medium
- 1
- Low
- 0
16 files · audited 2026-11-03 · rules 1.0.0
MediumAA-P02Can write files and run interpreters without disable-model-invocation
SKILL.md:4
Read Write Glob Grep Bash(node *)Why The agent may load this skill on its own and then write files and run code in one go.
Fix Set disable-model-invocation: true so the user starts it, or narrow the Bash patterns.
What this skill can use
Tools
- Read
- Write
- Glob
- Grep
- Bash(node *)
Scripts
- scripts/_lib.mjs
- scripts/doctor.mjs
- scripts/init.mjs
- scripts/pipeline.mjs
Domains mentioned
None
Automated checks found the patterns listed here. An audit reduces risk; it is not a guarantee of safety. How audits work · Read the rules
Show this grade in a README
The badge updates when the registry is rebuilt. Paste this Markdown into the README:
[](https://www.agentsaisle.dev/skills/video)SKILL.md
Rendered from commit 2ee11a4
Video
The user says what video they want in their own words; you make the best short the local tools allow. Three formats, each with a tested look, so every video starts from something good instead of a blank page. <skill> is the folder that holds this SKILL.md; the parts are its sibling video-... skills, which you load yourself. The user never names a part.
Use when
- The user wants a short video: explain an idea, algorithm, API or tool (explainer); punchy words on a voice, a hot take, a motivational or announcement short (kinetic); or a demo of their own app or site (showcase).
- The user asks what is possible, or why a tool is missing (run
doctor)./video initand/video doctorstill work as shortcuts.
Don't use when
- The request is for cartoons, characters, AI-generated footage, footage editing, dubbing or a long video. Say in one sentence that this skill makes three kinds of short (explainer, kinetic, showcase) and offer the nearest one.
- The user wants a video uploaded. The last step prepares a pack to upload by hand; nothing is ever uploaded.
- The request is for a known character, a real person's likeness, someone else's footage or logo, or a fact you cannot source. Say no in one sentence and offer an original alternative.
- The folder is a code project that should not get video folders. Ask for an empty folder first.
Steps
- Read the request and pick the format with
references/choose.md. Tell the user the plan in at most three short lines: the format, the length (default 30 to 45 s, vertical 9:16), and what you need from them. - First use in a folder:
node <skill>/scripts/doctor.mjs --cached; no studio yet:node <skill>/scripts/init.mjs. If the Remotion engine is missing, runnode <video-core>/scripts/engine.mjs setupand give the user the two lines it prints. Never install anything yourself and never read or ask for an API key. - Ask only what the request lacks (topic, audience, the app's address, brand colours). At most three questions, each with a default. If nobody is there to answer, take the defaults and say which.
- Read
references/craft.mdbefore writing a word: it is the quality bar every video is held to. Then follow the format's chain:references/explain.md,references/kinetic.mdorreferences/showcase.md. Runnode <skill>/scripts/pipeline.mjs projects/<slug>between steps to see what is next. - Pause for the user only at: the script and storyboard (once, before the voice), the demo script before a browser is driven, and the draft. Nothing else.
- Draft first (
render.mjs --draft), then review it asreferences/craft.mdsays: a contact sheet, the score card, and fixes until every score is 4 or more (at most 3 rounds). Use a separate reviewer agent when you can start one. - Render the final, run
video-qa, and look at its contact sheet. If a check fails, name it and the step that fixes it, then go back. - Offer the publish pack (
video-publish): show the MP4 path, the score card, the QA lines and the credits. Never upload. - Be honest about limits: synthetic voices are labelled, estimated word times can drift a little, and no one can promise a video will trend.
Budget
-
Search first (grep, glob, symbol search). Read line ranges, not whole files.
-
Read at most 8 files per step. If you need more, write down what you learned and narrow the question.
-
Do not re-read a file you just wrote or edited.
-
Prefer the smallest command output that answers the question: head, tail, quiet flags, filtered lists.
-
Run each script once per change. Do not read the other parts' files to describe them;
pipeline.mjssays what is next. -
Read
references/craft.mdonce per video, not once per scene.
Output
- The plan in three lines, then one line of progress per step.
- At the end: the video's path, its score card, its QA result and where the publish pack is.
Done when
- The draft scored 4 or more on every line of the score card, the final render passed
video-qa, and the user has the MP4 path. - Nothing was installed, downloaded, uploaded or overwritten without the user's yes, and every on-screen claim came from the user or a source in
facts.md.
Version history
| Version | Commit | Install |
|---|---|---|
| 0.29.0latest | 2ee11a4 | add video |