SteadyVane
Daily brief · October 13, 2026

Matt Pocock's agent skills workflow, six Claude repos and a free 32GB GPU setup

3 items that moved, with the numbers. Measured by SteadyVane, fact-checked, and graded before publishing.

YouTubeHow to run a coding agent through Matt Pocock's skills workflow, from spec to reviewPlay here

How to run a coding agent through Matt Pocock's skills workflow, from spec to review

YouTube · How-to · Ship like the best engineers: Matt Pocock's AI…: views history
47,767 → 128,107views in 28 h · 46 commentsabout 19× the typical pace of other tracked AI-tool videos

Matt Pocock's mattpocock/skills is an MIT-licensed GitHub repo of agent skills taken from his own .agents directory, built for coding agents such as Claude Code and Codex. The video follows one small feature through his flow of grilling, spec, tickets, implementation and review.

A video walking through it went from 47,767 to 128,107 views in 28 hours - about 25.9 times the typical pace of AI-tool videos SteadyVane tracks on YouTube. It has 1,809 likes and 46 comments, about 91 hours after it was published.

Why it matters: Agents can now build a feature unattended, so the open question is how you direct and check them. Pocock's answer is a reusable set of skills you install once, using 'npx skills@latest add mattpocock/skills', rather than a fresh prompt each time. Per his own X post, the repo now has more stars than React.

The agent writes the code. The skills decide whether it was asked the right question.
YouTubeHow to pick Claude add-ons: one creator's six best GitHub repos out of 100+ triedPlay here

How to pick Claude add-ons: one creator's six best GitHub repos out of 100+ tried

YouTube · How-to · I Tried 100+ Github Repos for Claude. These 6…: views history
139,971 → 156,204views in 12 h · 86 commentsabout 9.3× the typical pace of other tracked AI-tool videos

Nate Herk says he tried more than 100 GitHub repos built for Claude and picked six as the best. His community post says some made his projects 10 times faster and some replaced tools he was paying for monthly.

A video about it went from 139,971 to 156,204 views in 12 hours - about 11.9 times the typical pace of AI-tool videos SteadyVane tracks on YouTube. It has 3,272 likes and 86 comments, about 46 hours after it was published.

Why it matters: Sorting through Claude add-ons is a time cost for anyone building with it, and a shortlist saves some of it. The six repos are named only in the video, and the 'best' ranking is the creator's own judgment. The description also promotes a free skill, so part of the video is promotional.

A shortlist is a starting point. It is still one person's shortlist.
YouTubeHow to run a 27B coding model for free on Kaggle's two T4 GPUsPlay here

How to run a 27B coding model for free on Kaggle's two T4 GPUs

YouTube · How-to · Don't Buy a GPU Instead Build This 32GB Free AI…: views history
36,454 → 64,960views in 12 h · 144 commentsabout 16× the typical pace of other tracked AI-tool videos

Kaggle's free notebooks offer a T4 x2 option: two NVIDIA T4 GPUs with 16GB each, 32GB combined. A walkthrough shows a 27B model running on it through Ollama for local AI coding, and what the thumbnails leave out.

A video about it went from 36,454 to 64,960 views in 12 hours - about 18 times the typical pace of AI-tool videos SteadyVane tracks on YouTube. It has 1,503 likes and 144 comments, about 19 hours after it was published.

Why it matters: It is a no-cost way to try a large local model before buying hardware. The limits are real: about 30 hours of free GPU time a week, sessions capped at 12 hours, and two separate 16GB pools, so the model has to be split across both cards.

Free 32GB comes as two 16GB halves and a weekly quota. Fine for trying, not for living on.
← A tutorial has ChatGPT Codex edit 16 raw clips into a…

All briefs · What is rising now