// the find
msitarzewski/AGENT-ZERO
Operational framework and patterns for high-quality, AI-assisted software development. This repository provides a single canonical guide for AGENTS.md and a lightweight workflow for repeatable, auditable, and architecture-first delivery.
AGENT-ZERO is a prompt engineering framework — a single AGENTS.md file that tells AI coding assistants how to behave: approval gates, a state machine (PLAN→BUILD→DIFF→QA→APPROVAL→APPLY→DOCS), and a Memory Bank directory structure to persist project context across sessions. It is for teams who want their AI tools to stop freelancing and actually ask before applying changes. There is no code here, just instructions.
The state machine is well-thought-out — forcing a DIFF step before APPLY is the single most useful thing it adds over default AI behavior. The Memory Bank structure is opinionated enough to be useful without being so rigid it breaks on every real project. The compaction protocol (persisting state at every transition so context loss doesn't derail a session) solves a real problem that most teams hit but few address explicitly. The 'Four Sacred Rules' are pragmatic: reuse-first and cite-file-line are disciplines that actually improve output quality.
It's documentation pretending to be a tool — 259 stars for a Markdown file you have to read and hope your agent follows is a weak value proposition. Compliance is entirely trust-based; there's no tooling to enforce the gates or validate Memory Bank structure, so the whole thing collapses if the agent decides to skip PLAN and just BUILD. The Memory Bank pattern duplicates what a well-maintained README and ADR directory already give you, and if the team isn't already writing those, they're not going to maintain 14 separate context files either. Versioning at '2.2' with a date in the README is doing a lot of work to make a text file feel like software.