---
title: "My AGENTS.md file for building plans you actually read"
tags:
  - wiki
  - sources
sources:
  - "[[My AGENTS.md file for building plans you actually read]]"
domains: [agentic, coding]
date: "2026-07-17T06:23:09.809Z"
created: "2026-07-17T06:23:09.809Z"
includeInRss: false
---

# My AGENTS.md file for building plans you actually read

## Takeaways

- **Plan loop**: Plan → Execute → Test → Commit — skipping planning forces the AI to guess and produces hallucinations and rework.
- Planning forces **alignment before code**; the AI implements a agreed spec rather than inventing architecture mid-stream.
- Matt Pocock's plan-mode rules: make plans **extremely concise** (sacrifice grammar for brevity) and end with **unresolved questions** if any.
- Drop these rules into `AGENTS.md` or `CLAUDE.md` to make plan mode output scannable and actionable.
- Transforms AI from unreliable generator into **indispensable coding partner** when the loop is disciplined.
- Complements longer research/plan markdown workflows — this is the lightweight schema-level version.

## Notes

Minimal, copy-paste plan-mode conventions for agent config — directly applicable to this vault's `AGENTS.md` and Cursor rules.

## Open questions

- Should Ben merge Pocock's concision rules with Boris Tane's richer annotation cycle, or keep them as separate workflow tiers?
