Pocock's prompt rewrites AGENTS.md across three subagents
One prompt, three escalating rewrites of your agent instructions file, and one PR to compare them in.
Matt Pocock posted a prompt of the day aimed at the file most agent setups never clean up. It opens with his /writing-for-agents command, says the AGENTS.md is a hot mess, and asks for a series of restructurings that remove no-ops, use progressive disclosure, and move instructions into a CODING_STANDARDS.md.
Apply your work over three subagents, each more radical than the last. Create a single PR.
The interesting part is the last instruction. Rather than asking for one rewrite, the prompt tells the agent to apply the work over three subagents, each more radical than the last, and to create a single PR. You end up reviewing a spread of options in one place instead of accepting or rejecting a single answer.
Why the escalation matters
A single refactor prompt forces the model to pick an aggressiveness level on your behalf, and it usually guesses conservatively or guesses wrong. Three passes at increasing radicalism turn that guess into a choice you make at review time, which is the part of the loop a human is actually good at. The pattern is not specific to AGENTS.md. Any refactor where you would rather compare than commit can use the same final line.
Pocock's stated reason for targeting the file is cost. He calls AGENTS.md the most common source of token inefficiency, meaning instructions that get loaded into context on every run whether or not they do anything. No-ops are the lines that tell the agent something it was going to do anyway. Progressive disclosure, in this framing, means keeping the entry file short and pushing detail into files the agent reads only when it needs them, which is what the CODING_STANDARDS.md split is for.
This is a prompt, not a measurement. There is no before and after token count in the post, and no claim about how much the restructuring saves on any particular repo. What is on offer is a shape for the request.
It follows earlier prompts from Pocock in the same vein, short instructions that change what the agent produces rather than new tooling.