Skills For Real Engineers, published as the GitHub repository mattpocock/skills, is a collection of markdown instruction files that tell AI coding agents how to behave while they do software engineering work. It ships no application code at all: every file is prose that an agent reads before it starts a task. It is aimed at developers who already run Claude Code, GitHub Copilot or Codex and want the agent to follow a visible process instead of improvising one.
What it does
The pack is written by Matt Pocock, a TypeScript educator, and released under the MIT license. Each skill covers one piece of ordinary engineering practice:
- a test-driven-development loop, so the agent writes tests and keeps running them
- "grill-me", which makes the agent interview you about the requirements before it builds anything
- "wayfinder", for planning a large project before the first file is touched
- issue triage
- structured code review, including review of the agent's own output
- handoff notes, so a later agent session can pick up where the previous one stopped
The repository's saved README records 288.2k stars, 24.2k forks and 1.5k watchers at the moment it was captured, and both Pocock and the channel's own write-up say it reached number one on GitHub trending within a day. Those are the project's and the channel's claims rather than an independent measurement, and no separate GitHub figures were collected for this entry.
How it works
There is nothing to compile and no runtime. Installing copies the markdown into the place your agent looks for skills, and the agent loads the relevant file when the task matches. Because each skill is a separate small file, you can read it, edit it, delete the ones you dislike, or keep only the code-review skill and ignore the rest. That legibility is the design argument: Pocock positions small, composable, engineer-controlled instructions against one large autonomous agent that takes the human out of the loop.
The repository also ships a plugin manifest for Claude Code and GitHub Copilot, plus a one-time setup command that records your issue tracker and documentation preferences so the skills know where to file things.
Getting started
The repository itself is free, MIT-licensed and small, but it is not the part that costs money. You need a working coding agent first, which means Claude Code, GitHub Copilot or Codex, each with its own paid API key or subscription. The skills add no intelligence of their own, so whatever you already pay a model provider is what you keep paying. The project does not list hardware requirements, operating systems or memory, which is consistent with a pack of text files.
Install paths, as given in the README:
npx skills@latest add mattpocock/skillsclaude plugin install mattpocock-skills@claude-plugins-official- a GitHub Copilot plugin
Then run /setup-matt-pocock-skills once to configure the issue-tracker and docs preferences. There is no binary and no Releases page to visit: everything arrives through npm or your agent's plugin system, and you can also just read the markdown at github.com/mattpocock/skills.
When to use it / when not
The main catch is the one the video states plainly: this is instructions, not intelligence. The agent still does all of the work, and a skill file cannot make a weak model competent or stop it from producing something wrong. It changes the order of operations and raises the chance that tests get written and questions get asked.
Use it instead of just prompting Claude directly when you keep receiving code with no plan behind it and no tests around it, and you would rather spend five minutes answering the agent's questions than an hour reviewing a pile of output.
Don't use it if you want an agent that runs unattended from start to finish: these skills deliberately keep the engineer in the loop, which means more interruptions, not fewer. There is also nothing to use here until one of the supported agents is set up.
Alternatives
The alternative every reader already knows is vibe coding: open the agent, describe the feature, accept whatever comes back. That is precisely what this repository is positioned against. The other is a heavier end-to-end AI agent framework, which Pocock argues hides decisions from the engineer and costs you visibility and control. A third option is writing your own instruction files, since nothing here is more than markdown; the pack's value is that someone has already worked out what to put in them.
Take this repository seriously if you already pay for a coding agent and are tired of reviewing untested output from it. The cost of trying it is one command and some reading time, and the worst case is that you keep two skill files and delete the rest. If you do not use an agentic coding tool yet, there is nothing here to run: this is a process written down, for people who already have the tool and want some discipline around it.