Matt Pocock made a point recently that stuck with me: agents have no memory. They don't remember what worked last time. Without strict, well-defined processes, you're steering manually every single session.
That framing recontextualised a lot of frustrations I'd had with AI-assisted development. It's not that the agent is bad — it's that I was expecting it to carry forward context and judgment it simply doesn't have.
His Solution: Skills
Matt's answer is to build skills. Short SKILL.md files that give the agent a defined path to follow for a specific task. Instead of explaining what you want each time, the skill encodes the process once. The agent reads it and executes.
In his words, the code quality from AI shot up dramatically once he had skills in place. That tracks. A well-defined process produces consistent output. An open-ended prompt produces variable output.
The Skill Worth Calling Out
His post covers the five skills he reaches for every day. The one that caught my attention is /grill-me — it interviews you relentlessly about a plan before writing a single line of code. It's a forcing function for thinking through requirements properly rather than jumping straight to implementation. If you've ever shipped something and realised mid-way that the spec wasn't actually clear, this is the skill you needed first.
All of his skills are installable via:
npx skills@latest add mattpocock/skills/[skill-name]
Why This Matters
The pattern here — encoding process into files the agent reads — is the same instinct behind the Claude Code scaffolding system we wrote about last week. Knowledge bases, agent definitions, session logs: all of it is just giving the model a defined path rather than an open field.
Matt's framing makes it even cleaner: if the agent has no memory, your only lever is the quality of the process you hand it. Skills are that process, made reusable.
His full post and the GitHub repo are linked below — worth reading if you're using Claude Code seriously.
