Alvaro Cintas

Claude Cowork complete crash course from zero to expert in 12 minutes.

Installation, automations, connectors, plugins..all of it:

00:00 Intro
01:00 Installation & setup
02:08 Your first task
03:01 Global instructions (most skip this)
04:00 What Cowork can actually access
05:15 Skills feature
06:08 Connectors explained
07:04 Real workflow (Drive + Gmail)
07:46 Claude in Chrome
08:43 Scheduled tasks
09:52 Plugins (why $285B was wiped out)
11:41 Big picture

There is finally an open world model you can download and run on your own machine.

It's called LTX-2.5, from Lightricks, and it's built for film production, real-time experiences, and physical AI.

It comes with native 4K and HDR, native multi-shot scenes, and editing of your own real footage, on top of text-to-video, image-to-video, and audio-to-video generation.

The craziest part? 

The full 22B-parameter weights are open, so you can fine-tune it and run it on your own GPU, with no data leaving your machine. 

It's free for companies under $10M in annual revenue.

A 27B open-source model just took #1 on Artificial Analysis 🤯

Drex Max 1.1 from Nace AI picks between options across text, images, and video, and it's small enough to run on a Mac.

Runs on your own machine.

Small specialized models are winning.

You can now reverse-engineer any web animation into your own code.

It's called reanimate, a free Claude Code skill.

Ask an agent to rebuild a motion from a screenshot and it usually guesses once and stops.

reanimate films the original, rebuilds it in your stack, films the copy the same way, then compares timing, easing and pixels until the numbers match.

What you can feed it:

- A live website URL
- A screen recording
- An mp4 or gif
- A Lottie file

It writes into what your project already uses: CSS, Framer Motion, GSAP or three.js.

Usually 2 to 5 fix rounds. Whatever still doesn't match is listed in the report.

Install: npx skills add NiazMorshed2007/reanimate -g

Link: github.com/reanimate

A voice model just got small enough to run inside a browser tab.

It's called Woof, from Underdog. One voice is 31.7 MB. Two voices are 40.6 MB.

No server. Once it's loaded, you can go offline and it keeps talking. I set my browser offline and recorded it.

What makes it interesting is the hard text. Phone numbers, dates, prices and codes are where small voice models fall apart. Woof runs everything through a text normalizer first.

Underdog's own numbers on 500 hard sentences:

- Two-voice Woof: 1.42% of words wrong
- Fish Audio Drama 3: 1.89% wrong, and a 41.5 GB download
- One-voice Woof: 2.06% wrong

So it's about 1,000x smaller, with similar accuracy on their test.

Link: underdog.ai/woof