Key is letting the agents themselves write the rules. Have an agents.db file that lives in the repo root, with an extremely simple SQLite MCP tool that uses FTS5 search provided to all agents working on the code. The tool description instructs them how, when, and where to coordinate their changes and best practices for the repo, and the agents figure it out themselves from there. No complicated schemas or vector embedding, pure plain text in a single table. Any new agent can come in and figure it out immediately by exploring the DB, and they get temporal context. Any bad practices can be prompted against while running the agents, and they'll record it to the db in a way that makes sense to them. I do this for all of my projects. Markdown files aren't enough, and can't be read/written concurrently.

The research is:

Evaluating AGENTS.md: Are Repository-Level Context Files Helpful for Coding Agents?

https://arxiv.org/abs/2602.11988

---

Overall, it makes sense to me that excessive and/or unnecessary instructions dilute the importance of each instruction. Ideally it shouldn't, but it does.