Solo project
retained text
I’m building a hobby app and I’ve settled into a set-up that mostly works. I’d love tips from people running something similar. **How it runs now:** * **One session I talk to:** a “chief” on Opus. It takes decisions with me one at a time, writes briefs, checks proof and keeps the records. * **Workers:** * routine research goes to Sonnet agents; * fact-checks and corrections go to Opus agents; * a separate “second reader” agent re-opens the original sources before anything counts as approved. * **A Fable advisor** for hard calls (planning a study, auditing research, final reviews). It sends memos to the chief and never talks to me. * **Side sessions** for separate tracks (brand, a photo-recognition track, a study library, a build worker, trip planning). The chief briefs and messages them. * **Records after every task:** a log line, a one-page status card, a decision log, then commit and push. A stop hook nags if it’s skipped, and a git hook stops sessions committing each other’s files. * **Phone inbox:** I drop links into Notion, and a cloud routine writes a first look on each one every hour. **Where it hurts:** 1. **Usage limits.** Fable ran out mid-audit once. How do you decide what’s worth the top model? 2. **Long sessions.** The chief gets compacted and loses detail; once it couldn’t recall my own first prompt. Do you hand over to a fresh session, and how do you make sure nothing’s lost? 3. **Claude checking Claude.** An outside human reviewer caught things the agents missed. What’s your cheapest reliable check? 4. **Too many threads.** I keep starting side sessions. How many do you run at once before it gets messy? 5. **Rules that slip.** Written rules in [CLAUDE.md](http://CLAUDE.md) sometimes get skipped. What do you enforce with hooks instead? What would you cut, and what’s missing?
document fetched 2026-11-12 00:43 · canonical https://reddit.com/r/ClaudeCode/comments/1xt6666/solo_project