I've coded with AI for 2 years and this is my whole setup
retained text
I keep seeing people ask what setup to use and getting answers like "just use claude code bro". It works until the project gets bigger than a weekend. So here's what I actually run day to day after 2 years. It's not perfect, it's just what survived. Coding: claude code for anything that needs a plan, codex for the boring background stuff I don't want to babysit. I stopped switching models every week, it cost me more time than it saved. Planning: I write the ticket in plain English before the agent touches anything. Three or four lines of what done looks like. Half my bad sessions came from skipping this. Rules: one AGENTS.md, under 100 lines. Mine was 900 lines at one point and the agent ignored most of it anyway. Review: coderabbit on every PR, plus a claude review pass locally before I push. Coderabbit was noisy the first week until I told it to skip style nits, now it's mostly real stuff. I still read anything touching payments or auth myself. Testing: playwright for the 5 flows that make money (signup, login, checkout, invite, cancel). The agent writes the rest of the tests and I don't trust them much. CI and deploy: github actions runs everything, vercel for the app, a $12 VPS for the worker. Errors: sentry. The agent can read the stack trace better than I can at 1am. What I dropped: 11 MCP servers down to 2, multi-agent setups that looked cool on twitter, and long sessions. New task, new session. well, nothing fancy, most of the gains came from cutting things, not adding them, at least for me
document fetched 2026-10-21 00:18 · canonical https://reddit.com/r/AI_Agents/comments/1xawuwy/ive_coded_with_ai_for_2_years_and_this_is_my