Every Session
- Start Right
- During Work
- Before Ending
- Run
/doctorif things feel off - Check context with
/context— know your budget - Name your session with
/renamefor easy/resumelater - For returning tasks:
/replay <topic>to surface past learnings
Every Day
Update
Terminal Setup
Use a Capable Terminal
- iTerm2, Ghostty, Warp, or Kitty
- Avoid IDE embedded terminals for heavy Claude Code sessions
- tmux for split pane agent teams
- Short alias:
alias c='claude'
Monitor Usage
/usageto check plan limits/costfor API key users- Watch for fast mode billing (charges extra usage from first token)
Every Week
Review Patterns
Identify Hot Learnings
Identify Hot Learnings
Corrected often, but not yet learned. These should become rules.
Prune Cold Learnings
Prune Cold Learnings
Learned but never applied. Remove to reduce context noise.
Optimize CLAUDE.md
- Keep root CLAUDE.md under 150 lines
- Move specialized knowledge to package-level CLAUDE.md files
- Move personal preferences to CLAUDE.local.md (gitignored)
Debugging Tips
- When Claude Gets Stuck
- When Context Is Degraded
- When You're Blocked
- Try a different model — switch to Opus for hard problems
- Use
ultrathinkin your prompt for maximum reasoning - Provide a screenshot (Claude is multimodal)
- Share browser console logs via MCP
- Start a fresh session with more targeted context
Environment Tips
Terminal Recommendations
| Terminal | Platform | Why |
|---|---|---|
| iTerm2 | macOS | Split panes, tmux integration |
| Ghostty | macOS/Linux | Fast, GPU-rendered |
| Warp | macOS | AI-native, blocks |
| Kitty | Cross-platform | GPU-rendered, scriptable |
| Windows Terminal | Windows | WSL support |
Voice Prompting
If you have voice mode access:Tool Layering
The most productive setup uses multiple tools:Anti-Patterns
| Don’t | Do Instead |
|---|---|
| Skip quality gates to save time | Gates prevent more corrections later |
Use dontAsk permission mode daily | Use default with specific allow rules |
| Load 15+ MCP servers | Keep <10 active, disable unused |
| Compact mid-task | Compact at task boundaries |
| Fight the AI on style | Add rules to CLAUDE.md instead |
| Use one huge CLAUDE.md | Split into root + package-level files |
| Ignore the correction pattern | Every correction is a learning opportunity |