Your Repo Has To Be the Boss
The most useful AI coding pattern I saw this week was simple. The people getting steady results were not relying on hidden memory or one hero model. They were writing the rules into the repo, forcing the agent to show its plan, and using a different model family to review the result. That sounds less glamorous than "god mode", but it is much closer to the setup I would trust on a real product.
Stop hiding the rules
One of the better Codex threads this week made the case for setting up AGENTS.md before anything else. That landed for me because it matches the part that actually matters in practice. If the repo does not say how the product is structured, what the conventions are, and where the agent has room to act on its own, the model is guessing. Sometimes it guesses well. Sometimes it just writes something plausible enough to pass a quick skim.
A related Claude Code thread pushed in the same direction from the other side. People were actively asking whether project context belongs in memory or in visible files. The good answer was the boring one. Put project truth in files you can grep, diff, and review. Use memory for personal preferences or machine-specific context if you need it, but do not let the repo lose its job as the source of truth.
Drift is the real tax
The thread I kept coming back to was from a builder who got tired of switching between coding agents and watching the codebase slowly drift. Not explode. Drift. Test names changed shape. Folders started bending. Imports moved in ways that still compiled but made the architecture worse. That is the kind of damage that wastes a week without announcing itself.
What I liked about that post was the framing. The fix was not "use a smarter model." The fix was to write down standards explicitly and check every PR against them. That is a much more honest view of the problem. If I am moving between Codex, Claude, and whatever else is useful that day, I need the standards to be stable even when the agent is not.
One model should not grade its own paper
The loudest workflow thread this week was the so-called God Mode setup, where Claude, Codex, and other model families review the same work in parallel. The part worth keeping is not the theater of running a whole control room. It is the simple observation that same-family models tend to share blind spots, while a fresh reviewer from a different lineage catches bugs and design drift that the first one glides past.
That feels immediately useful to me. I do not need a giant orchestration rig to borrow the lesson. I can keep one model as the main implementer, then run a second one as a skeptical reviewer on plans, diffs, or risky files. The point is not to admire consensus. The point is to force disagreement before the code lands.
The setup I would actually keep
If I were tightening my own workflow around this, I would keep it pretty small. One repo-level instructions file. One standards file for naming, architecture boundaries, and testing expectations. Mandatory plan review before big edits. Then one reviewer from a different model family for anything broad, risky, or expensive to unwind later.
I would also be strict about what does not belong in hidden state. If a rule matters to the product, it belongs in version control. If a preference matters only to me on my machine, that can live outside the repo. That boundary seems dull, but it is exactly the sort of dull rule that keeps a tool useful once the novelty wears off.
Why this was worth writing about
I wrote this one down because it feels like a real upgrade in how builders are thinking. The conversation is getting less mystical. People are starting to treat AI coding like an engineering system with documents, checkpoints, and failure modes instead of a magic coworker that hopefully remembers what you meant.
From where I sit behind Applikeable, that is a healthy shift. Small teams do not get much value from fancy orchestration for its own sake. We get value from clearer rules, better review pressure, and fewer quiet mistakes making it into the product. This week had a lot of noise, but that lesson was worth keeping.