Matt Pocock on coding agents in the AI era: Say clearly what you care about, and make your codebase worthy of the agent

Matt Pocock: The most important skill in the AI era is being clear about what you care about, and making your codebase worthy of an Agent

From @nateherk's 44-minute interview with @mattpocockuk, Pocock believes: the most valuable thing is definitely not chasing the latest model, but two things: being clear about what you care about, and getting testing, code quality, and processes to the point where agents can be relied on; he himself tried spec-driven development for a few months, and his code actually got worse and worse.

Youtube video: https://t.co/jqPorXoisJ

When Pocock started to change his mind About ten years of self-teaching and about eight years professionally, he had always been rather skeptical of coding agents. Around December 2025, after Opus 4.5 came out, he felt the threshold had been crossed: he no longer writes a single line of code by hand. He had always loved improving his colleagues' working environment and teaching; once agents became good enough, the same skills shifted toward "how to make agents work better."

Core: tell the agent what you care about There is an often-underestimated communication gap between people and agents. It doesn't know what you actually want, and it doesn't know what you care about, so you have to explain it clearly over and over. If the codebase is unmaintained and all just cobbled together, the agent can only hand in garbage too—still garbage in, garbage out. The right approach is like covering for the worst human colleague: tests, checks, blocking a whole class of errors. Classic software engineering thus becomes useful again. Agents can teach you too, but stay skeptical and prioritize reading primary sources. His /teach finds trustworthy sources and cites them, using AI as a guide and then reading the original text.

Where the opportunities are Big changes, many opportunities: teaching people, building AI-embedded products, helping companies overhaul their systems. Consulting or internal promotion, it's the same—bet on skills you can take with you. No matter how strong models get, you still need to learn: how to trust, how to organize multiple agents, what information to give them, how to translate company processes into workflows agents can run. Code and math are easy to verify; real knowledge work is harder, and packaging this kind of process for organizations is a big opportunity.

Unknown unknowns and Grill Me "The Pragmatic Programmer" says: nobody knows what they want at the start. His high-frequency skill grill me has the agent keep asking questions to dig out the context you really care about; Grill with Docs then helps you build domain terminology (the ubiquitous language in DDD: code, database tables, UI, customers, and users all speak the same set of words). The key ability is realizing "there are still blanks you don't know about"; a design isn't complete until it's turned into code.

How to seem reliable "Weekend portfolio stacking" makes it hard to build trust. Clients want someone who understands what they care about. Using agents to quickly build a prototype that proves you understand the business is enough. The people who really do well are often the best communicators; once you've thought it through, PRs, products, and instructions to agents all become clear.

Why spec-driven development goes wrong (26:03) Inspired by "specs are the new code" from the OpenAI side at AI Engineer, he tried it for a few months in his personal Descript video-editing repo: the more detailed the specs got, the more he had the agent generate code. As a result, code quality slid all the way down, and the agent got worse and worse. He later concluded: the code itself is the agent's working environment; maintainable code that can be safely changed is more useful than living in specs all the time. Models can't read minds, skills and processes will be needed long-term; people also underestimate the harness (things like Claude Code and Codex)—don't just look at the model brand.

Models, Harness, Effort He now commonly uses Claude Code + Opus 5.5. One sharp view: most people set effort too high. He himself uses medium, and is considering going lower for faster responses and fewer tokens, because tests have already raised the floor. Letting the agent open a browser and look at screenshots is often more worthwhile than switching to a newer model; an old Opus 4.5 in a good environment can beat a new model in a bad environment. For the harness, Claude Code, Codex, or Copilot are all fine—don't agonize over it too long.

How to start First install a mainstream harness; whether to use his skills is up to you. He mentions the trust ladder from Lauren Tan (Potato on Twitter, who makes Grok Bot): learn to brake first, then step on the gas. An empty folder, one real idea, and watch it closely; then expand to multiple agents, custom skills, and hooking up Slack and email. Don't throw the hardest work at it right away; start with trivial, boring, measurable tasks where mistakes are cheap to fix (like drafting replies), then talk ambition.

What he's excited about next He considers himself a teacher, not a critic. He's thinking about software factories: how to reliably turn ideas into work and long-term changes to the repo; how to organize multiple agents still has no answer. He also looks forward to open-source models approaching Opus 5.5, and to models that use fewer tokens, speak more plainly, and can get work done.