Two engineers. Same weekly Claude Code limit. One burns out by Wednesday. The other ships all week.
The gap isn’t skill. It’s one config file.
I spent the weekend testing every hidden setting in Claude Code. Eight of them, each one a single line in settings.json, change how the whole week feels:
1. Plan on Opus. Build on Sonnet. Same output, roughly half the usage.
2. Stop maxing the effort dial. Independent timed tests (20 rounds/level) showed correctness doesn’t change below max. Extra tokens just buy re-verification and caveats.
3. Add deny rules first. Block .env reads. Block unattended git push. Deny is checked before allow.
4. Compact at 75%, not 95%. Quality degrades well before the default threshold.
5. Set cleanupPeriodDays to 365. The 30-day default is silently deleting your decisions.
6. Run /fewer-permission-prompts. It builds a tailored allow-list from your own history.
7. Turn on local sound + mobile push. Know the moment Claude is waiting on you.
8. /statusline shows model + context %. So you compact before auto-compaction does.
The meta-shift for 2026: stop writing longer hyper-specific prompts. Give the model the outcome, the context, and what “done” looks like. Then let it act. We don’t call it “prompt engineering” anymore. We call it loop engineering.
Save this. Send it to someone who hits the weekly limit too fast.
#ClaudeCode #AIEngineering #DevTools #DeveloperProductivity #AISettings #ClaudeAI #Anthropic #LLM #ProductivityTips #EngineeringLeadership
— @devengoratela
Video Source
About the Author
