XNews

Matt Pocock Says Clear Intent and Reliable Codebases Matter More Than New AI Models

Matt Pocock Says Clear Intent and Reliable Codebases Matter More Than New AI Models

In a 44-minute interview with Nate Herk, Matt Pocock argued that the most valuable AI-era skills are articulating what you care about and making tests, code quality and processes reliable enough for agents, rather than chasing the newest models. He said months of spec-driven development in his Descript video-editing repository produced steadily worse code, concluding that maintainable code is the agent's working environment. Pocock also said he stopped hand-writing code after Opus 4.5 and now uses Claude Code with Opus 5.5 at medium effort.

Event version 1.

Key insights

Communication gap with agents

Pocock said the gap between people and agents is often underestimated because agents do not know what users want or care about. He compared the correct approach to covering for the worst human colleague: tests, checks and blocking classes of errors, which makes classic software engineering useful again.

Spec-driven development backfired

Inspired by the idea that specs are the new code, Pocock tried increasingly detailed specs feeding agent-generated code in his Descript video-editing repository for several months. He said code quality declined and the agent performed worse, leading him to conclude that maintainable code is the agent's working environment.

Effort settings and harness choice

Pocock said most people set effort too high and that he uses medium, considering going lower for faster responses and fewer tokens because tests raise the floor. He said giving agents browser access and screenshots often pays off more than a newer model, and that harness choice among Claude Code, Codex or Copilot should not be agonized over.

Grill Me and domain language

Pocock's high-frequency grill me skill has the agent keep asking questions to surface the context users actually care about. Grill with Docs builds domain terminology so code, database tables, interfaces, customers and users share one vocabulary, and he said design is not complete before it becomes code.

Starting with a trust ladder

Pocock cited Lauren Tan's trust ladder: learn to brake before accelerating, starting with an empty folder and one real idea under supervision. He advised giving agents trivial, boring, measurable, low-cost tasks such as reply drafts before expanding to multiple agents, custom skills and Slack or email integrations.

Source posts