1. Don't watch it work. Give CC a clear task with success criteria, then go do something else, directory submissions, Quora answers, whatever. Come back to finished work or a clear error.
2. Batch related work into one session. Instead of "fix this button" then "add this feature" - give it a full plan with 6-8 items. It enters plan mode, works through them sequentially, and you review once at the end.
3. CLAUDE.md with workflow rules. CC reads it at session start. Mine includes "never mark a task complete without proving it works" and "after any correction, update lessons.md." This eliminated the cycle of fix break ask fix.
4. Accept the async mental model. Agentic coding isn't pair programming. It's more like delegating to a junior who's incredibly fast but needs clear specs. Write the spec well, step away, review the output.
The broken rhythm only hurts when you try to synchronously watch and steer. The moment I started treating CC sessions like async tasks, everything clicked.
Just implemented Boris Cherny's (engineer behind Claude Code) CLAUDE.md workflow in my project. Game changer.
Key rules that made the biggest difference:
1. Plan Mode Default — enter plan mode for any task with 3+ steps. If something goes sideways, STOP and re-plan instead of pushing forward.
2. Self-Improvement Loop — after ANY correction, CC updates tasks/lessons.md with the pattern. It stops making the same mistake twice.
3. Verification Before Done — never mark complete without proving it works. "Would a staff engineer approve this?"
6. Autonomous Bug Fixing — when given a bug report, just fix it. Don't ask for hand-holding. Zero context switching from the user.
Before this, CC would ask permission for everything and repeat mistakes across sessions. Now it plans first, fixes autonomously, and learns from corrections.
The CLAUDE.md file persists across sessions — it's basically giving CC a working memory and engineering standards in one file.
I use Claude Code for basically everything around my product, not just the code itself:
- SEO audits (checking structured data, hreflang across 25 language versions)
- Batch blog translations (40 articles × 25 languages)
- Performance debugging (found that our entire SSR body was empty, LazyAuthProvider with ssr:false was wrapping all content)
- Admin dashboard analytics queries
The biggest unlock was giving it a good CLAUDE.md with workflow rules. It stops asking permission and starts fixing things autonomously.
Building a browser-based file tools product and CC handles probably 80% of the iteration cycle at this point.
1. Don't watch it work. Give CC a clear task with success criteria, then go do something else, directory submissions, Quora answers, whatever. Come back to finished work or a clear error.
2. Batch related work into one session. Instead of "fix this button" then "add this feature" - give it a full plan with 6-8 items. It enters plan mode, works through them sequentially, and you review once at the end.
3. CLAUDE.md with workflow rules. CC reads it at session start. Mine includes "never mark a task complete without proving it works" and "after any correction, update lessons.md." This eliminated the cycle of fix break ask fix.
4. Accept the async mental model. Agentic coding isn't pair programming. It's more like delegating to a junior who's incredibly fast but needs clear specs. Write the spec well, step away, review the output.
The broken rhythm only hurts when you try to synchronously watch and steer. The moment I started treating CC sessions like async tasks, everything clicked.