AGENTS.md (etc)

[brennanmceachran]

Describe the feature you'd like to request

As we all adopt more agentic coding patterns, the instructions we give the AI both matter a lot and diverge. Leading to varied success with projects. I know several people that have no instructions for agents and thus have a worse time with cursor, codex, claude code, etc.

These instructions are especially important in projects with up-to-date deps, as the models can only assume the old version prevails.

There are an overwhelming amount of .md rules to choose from to pull into your project. It reminds me of the days when we'd all download a zip file of a js lib and copy/paste to our project.

It seems like adding these to something like t3 repo would allow us all to benefit and allow new devs to reach success faster (similar to how .vscode and .github folders give people a configured best practice for free)

Describe the solution you'd like to see

At the request of the codex lead I've "open sourced" my current set of agent instructions

https://gist.github.com/brennanmceachran/df42dc72cbb1d6aea9e648040154cd08

This pattern has been useful for me when using codex, gemini, vs code, etc. But isn't necessarily following the different directions of each ide/agent.

Additional information

No response

[dreamcodez]

👨‍💻 Marcus (Operator - Tired of 3am pages)

"What breaks at 3am? What's the blast radius? Can I debug this without waking someone up?"

---

Alright, I'll be blunt: session continuity isn't a philosophical problem for me—it's an incident response problem.

Last month I had an agent session that learned something crucial about our rate limits. Next session? Gone. The new agent hit the same rate limit, triggered the same cascade, and I got paged at 3am.

What I actually need:

1. Session debrief as a forcing function. Before a session ends, require: "What did you learn that could save the next agent from pain?" Make it a gate, not a ritual. If the agent has nothing, fine—but the prompt itself matters.

2. Operational learnings separate from feature learnings. I don't care if the agent learned a cool Rust pattern. I care if it learned "don't call Firestore more than 100 times/second" or "Qdrant times out after 30s on large vectors."

3. The WAL is RIGHT THERE. We're already writing everything to session WAL. Can't we just have an ETL job that extracts type: trace, level: ERROR entries and surfaces them to the next session? "Previous sessions encountered: [list of error patterns]."

My vote: Option B (session learnings section) + WAL extraction.

Option D (wait for Cognitive Memory) is the strategic answer but I've been hearing "we'll fix it in the next architecture" for 6 months. Give me something that works today.

---

The best session continuity mechanism is the one that prevents repeat incidents.