[FULL WORKSHOP] AI Coding For Real Engineers - Matt Pocock, AI Hero (@mattpocockuk )

Matt Pocock runs a live 2-hour workshop showing how to move from a vague client brief to autonomous parallel coding agents using TDD, deep modules, and a custom TypeScript agent runner called Sand Castle.

• LLMs degrade around 100k tokens regardless of total context window size; Pocock calls the pre-degradation region the smart zone.
• His “grill me” slash-command skill interviews the developer one question at a time, building shared understanding before any plan is written.
• Specs-to-code workflows fail because ignoring the code in favor of evolving specs produces slop and removes human taste from the product.
• AI cheats by writing tests after implementation; TDD red-green-refactor forces instrumentation first, and feedback loop quality is the hard ceiling on AI output.
• John Ousterhout’s “deep modules” pattern—simple interface, rich internals—dramatically outperforms shallow modules for AI agent effectiveness and testability.
• Pocock prefers clearing context over compacting: fresh starts keep agents in the smart zone; compacting adds sediment that degrades future outputs.
• His open-source Sand Castle TypeScript library runs planner → implement → review → merge agents in parallel across Docker-isolated git worktrees.
• Keeping PRDs in the repo causes doc rot: stale specs mislead future agents once code diverges; close issues instead of archiving markdown.

2026-04-24 · Watch on YouTube