Matt Pocock scales AI agents

Matt Pocock is pushing a structured way to scale coding agents: plan first, break work into slices, then let agents execute in loops. (youtube.com) The workflow starts before code. Pocock’s `grill-me` skill tells the model to interrogate a plan until there is “shared understanding” and to inspect the codebase when answers already exist there. (github.com) (aihero.dev) From there, the plan becomes a Product Requirements Document, or PRD, and then a set of independently grabbable issues. Pocock’s public `skills` repository lists `to-prd` and `to-issues` as separate steps, with `to-issues` explicitly aimed at vertical slices. (github.com) (youtube.com) That structure is built around a limit Pocock keeps returning to: large language models do better on bounded tasks than sprawling ones. A workshop summary tied to his April 24 session says he frames fresh context windows of roughly 100,000 tokens as the model’s “smart zone.” (github.com) (youtube.com) Once the work is sliced, Pocock’s “Ralph” loop turns the agent into a worker that repeatedly reads a PRD, finds the next incomplete task, implements one item, commits it, and updates progress. His guide says the setup uses Claude Code plus Docker Desktop. (aihero.dev) Docker is there for containment, not convenience. Pocock’s guide says the sandbox lets the agent run commands, install packages and modify files without touching the developer’s local machine. (aihero.dev) The public repo shows the rest of the guardrails. Alongside planning skills, Pocock includes `tdd`, `qa`, `git-guardrails-claude-code`, `github-triage` and `improve-codebase-architecture`, which turns the method into a software process, not a single prompt. (github.com) That emphasis matches Pocock’s broader pitch on AI Hero and YouTube: software engineering habits that worked for humans still work for agents. Planning, testing, modular design and review are the recurring controls. (mattpocock.com) (youtube.com) The result is less “ask the bot for code” than “run a managed assembly line.” Pocock’s own materials describe agents as useful but forgetful, which is why the documents, slices and QA steps stay in the loop. (aihero.dev) (github.com) Pocock’s system does not promise that agents fix bad specs or messy repositories. It assumes the opposite: the more disciplined the plan and the codebase, the more safely multiple agents can ship work while a human stays in charge. (youtube.com) (github.com)