# skills
A growing pile of [Claude Code](https://claude.com/claude-code) and [Codex](https://developers.openai.com/codex/) skills: drop-in folders that teach your agent to do one thing really, really well.
Right now the pile has six skills in it, and more are coming.
> ⭐ **If this saves you an afternoon, leave a star.** It costs nothing and helps other people find the repo.
## The skills
| Skill | What it does |
|---|---|
| [`motion-designer`](./motion-designer) | Turns a product website, or any story or topic, into short motion-graphics videos for Reels, TikTok, Shorts, Instagram, LinkedIn and X. You get real photos, real 3D, captions that appear word by word as they're spoken, natural narration, sound design and a spoken call to action. Everything is rendered from code: no After Effects, no timeline scrubbing. Each video is checked frame by frame before you see it. Before it starts, it asks you a few questions: language, idea, how many versions, sound, voice, platform and length. |
| [`diretor-de-edicao`](./diretor-de-edicao) | Edits the short videos you shoot (Reels, TikTok, Shorts, YouTube, LinkedIn): word-level cuts with jump cuts, the "yap" format (you talking to the camera while text, real screenshots, logos and images pop up around you, never over your face), screen + face tutorials, word-by-word captions and viral sound effects at the right beats. Before editing it searches the web in the browser for the references, screenshots and images that prove what you say (Pinterest and image banks for style, primary sources for facts) and logs every source. Also writes the script with the rules that kept videos watched (no punchline filler, explain before the example, the first fact in 14 s) and tells you what to record. Starts with a short intake: language, what you have, format, platform. |
| [`webscraping-expert`](./webscraping-expert) | The whole playbook for getting data off a website or API. Starts with a short intake, then always tries the cheapest path first — a plain HTTP request before a browser — and only climbs the ladder when it has to: TLS/HTTP fingerprint impersonation, the modern stealth-browser stack, the anti-bot vendor matrix (Cloudflare, DataDome, Akamai, Kasada) with a bypass per vendor, CAPTCHA solving, proxy networks and managed unblockers with geo-targeting, and structured data (JSON-LD) the site hands you for free. Includes site-stack playbooks (Next.js/RSC, Tableau, OData, GraphQL, ASP.NET, SPAs and more), HAR-driven discovery, document extraction (PDF/XLSX/OCR), resumable runs, and a neutral recon report at the end. |
| [`agentbox`](./agentbox) | Runs web-scraping subagents inside isolated Docker sandboxes, one box per subagent. Each box has a terminal, a persistent headless Chromium, passive listeners that log every request without touching the traffic, and an optional proxy from any provider with a hard budget and known cost per request. An agent journal and a merged timeline let you check every claim a subagent makes before you trust it. Needs Docker. |
| [`aula-interativa`](./aula-interativa) | Turns any topic into a self-contained HTML lesson with interactive slides, worked examples and deterministic animations checked frame by frame with Playwright. Asks only for language, topic and depth (or a slide count). Includes detailed teaching and motion guides, ten animation recipes, a reusable player, frame capture, contact sheets and artifact-bound review reports. Software lessons connect concepts to version-labelled UI walkthroughs. |
| [`ui-humanizer`](./ui-humanizer) | Takes the "made with AI" look out of sites, landing pages, apps, docs and slides. Built from 51 replies to a thread on what gives an AI-made site away: the uppercase label above every heading, the gradient word in the title, cards with a side stripe, pulsing chips, emoji as icons, Tailwind purple, pills everywhere, a bakery that looks like a SaaS. It runs a full process (brief, sketch before code, references from Pinterest and friends searched in the browser, design system, build, audit) and ships a dependency-free detector script. |
*Some skills are on their way. Some are still just a strongly worded note on my phone.*
## Install a skill
```bash
git clone https://github.com/gabrielgreco1/skills.git
cp -r skills/motion-designer ~/.claude/skills/
```
Or with the [skills CLI](https://skills.sh): `npx skills add gabrielgreco1/skills --skill agentbox`.
Then open Claude Code and just ask, e.g. *"make a motion video about how coffee gets from the farm to your cup"*, or call it with `/motion-designer`. Each skill's own README lists what you need to install for it.
For interactive lessons, install `aula-interativa` instead. Its [README](./aula-interativa/README.md) includes Claude Code and Codex setup.
## FAQ
**Does motion-designer really make the video, or just a plan for one?**
It makes the actual MP4s, more than one version if you want, and opens them in Finder so you can pick a favourite.
**Will motion-designer burn my CPU?**
For a little while, yes. 240 fps renders are no joke. It runs at low priority and cleans up after itself, so your laptop will survive. Your fans will have opinions.
**Can I contribute a skill?**
Yes. Open a PR with a folder that contains a `SKILL.md`. Bonus points if every rule in it came from a mistake you made yourself.
## License
Use the skills, change them, ship videos with them. The vendored third-party code keeps its own license (Three.js is MIT).
