๐ญ๐ฌ ๐๐น๐ฎ๐๐ฑ๐ฒ ๐๐ผ๐ฑ๐ฒ ๐ฆ๐ฒ๐๐๐ถ๐ป๐ด๐ ๐๐ผ๐ฟ ๐๐ฒ๐๐๐ฒ๐ฟ ๐ช๐ผ๐ฟ๐ธ๐ณ๐น๐ผ๐๐
I used Claude Code for six months. Defaults cost me 15 hours a month. The agent forgot decisions. It rewrote code without asking.
Here are the settings you need.
Keep CLAUDE.md under 8,000 characters. Long documents confuse the agent. Move rules to a folder structure:
- .claude/skills/ for session rules
- .claude/specs/ for domain models
Fix your settings.json. Use allow and deny lists for bash commands. This reduces interruptions.
Use acceptEdits for simple fixes. Use it for renames and import swaps. Turn it off for new features or data layer changes.
Set up hooks.
- Stop hook: Log session time.
- PreToolUse hook: Block commands like rm -rf.
- Notification hook: Get a ping when done.
Run complex tasks on three models. Use Opus, Sonnet, and Haiku. Use a final Opus session to pick the best path.
Set effort levels.
- Low: fast draft.
- Medium: short plan.
- High: detailed plan.
- Max: analyze and approve each step.
Restart long sessions. If the agent repeats a mistake, stop. Use a short handoff note to start fresh.
Reframe tasks after two failures.
- Split the task.
- Give an example.
- Write the skeleton first.
Use layered settings. Keep strict rules in the team file. Put personal preferences in a local file.
Use skills files. Put project context in .claude/skills/. This lowers token use and keeps the agent focused.
Claude Code is a platform. Configure it to move fast.
Source: https://dev.to/evan-dong/-10-claude-code-settings-i-wish-id-configured-on-day-one-3f57 Optional learning community: https://t.me/GyaanSetuAi