Intel / article

AI Skills with Matt Pocock

Source: newsletter.pragmaticengineer.com
Author: Gergely Orosz
Date: Published Sep 17, 2026

Context

Software educator Matt Pocock describes using a small, working path through an application to guide an AI coding . In the Pragmatic Engineer episode summary, Pocock says agents tend to build software layer by layer, which leaves bugs where the layers connect. After reading about the tracer bullet in The Pragmatic Programmer, a working path through the system, Pocock instructed an agent to build with tracer bullets and reports that it started to produce better code. The instruction changes the unit of work: the agent must connect enough pieces to make one behavior function before expanding the application. It illustrates how an engineering method can become a useful agent instruction, with the improvement presented as Pocock's experience, not a measured result.

Key takeaways · AI-distilled

The 'grill-me' was inspired by Anthropic's Thariq Shihipar, who found having an agent interview its user about a topic surprisingly useful. Pocock turned that into a short skill that instructs the agent to interview the user relentlessly before acting.
Pocock noticed agents building software layer by layer produced bugs between layers. Borrowing 'tracer bullets' from The Pragmatic Programmer (implement a golden path first) improved agent output, so he now mines classic software books for other 'leading words' that steer agents.
He frames codebase health for agents as 'Memento-driven development': like the amnesiac in that film, an agent starts fresh every session with no accumulated tacit knowledge, so the readability and structure fundamentals matter more for agent-maintained code, not less.
Pocock is moving his coding work to cloud-based agents over local ones because cloud agents keep running after he closes his laptop and can be made 'multiplayer,' letting multiple people collaborate on the same agent session in ways local setups can't.
He has mixed feelings on TDD with agents: failing tests help humans because of short working memory, but agents have longer windows, so he now asks agents to produce direct proof their code works, with or without a formal TDD loop.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
agent skill — A reusable instruction file that teaches an agent how to do one job well — the procedure, the tools, and what counts as done.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.

Read the source newsletter.pragmaticengineer.com