How we pair-program with Claude to ship daily

The setup

Our dev loop is Claude in the IDE + a small team of humans. The vast majority of code we ship is co-authored — the commit messages list both us and the model. Most days we land 4-8 PRs.

This is sustainable because the model handles 80% of the "what does the syntax look like, what's the standard pattern for this" overhead. We handle product judgment, architecture, debugging when it's stuck, and writing the failure cases.

What works

1. Scope each prompt narrowly. Asking Claude to "build the factory" failed early on. Asking for "the cancel button endpoint, given these existing types" lands a working PR every time.

2. Hand-pick which files the model sees. Don't dump the whole repo. Give it the 3-5 files it needs to know about. The output is 3x better.

3. Make it test its own work. Ask for the test next to the implementation. Even if you don't keep all the tests, they catch regressions immediately.

4. Read every diff. We don't auto-accept anything. Every PR gets a human review even if the model wrote it. That's where mistakes get caught.

5. Ship tiny, often. A 200-line PR that lands in 20 minutes is better than a 2000-line PR that argues for 3 hours.

What doesn't work

1. "Build me X" prompts that span multiple files. Always fails. We split them ourselves.

2. Leaning on the model for architecture. It's good at "implement this function," bad at "design this system." We do the system, it does the code.

3. Trusting it on unfamiliar APIs. It hallucinates. We always verify against actual docs.

How this affects you

When you use OOretz Factory, the same loop applies — at the orchestrator level. The factory scopes each generation step to one entity or one workflow. It hand-picks the files it shows the AI. It validates output via the verification loop. It catches mistakes via the audit log + AI confidence scoring.

We built the factory the way we wish AI tools were built for us.