Skip to content

daily-habits.md

Latest commit

 

History

History
122 lines (96 loc) · 4.08 KB

daily-habits.md

File metadata and controls

122 lines (96 loc) · 4.08 KB

Daily Habits for AI-Assisted Development

Practical habits organized by frequency. Tested across Claude Code, Cursor, and Codex workflows.

Every Session

Start Right

  • Run /doctor if things feel off
  • Check context with /context — know your budget
  • Name your session with /rename for easy /resume later
  • For returning tasks: /replay <topic> to surface past learnings

During Work

  • Commit early, commit often — as soon as a logical unit is done
  • Manual /compact at 50% — don't wait for auto-compact at 95% (the "dumb zone")
  • Read before edit — always understand what's there before changing it
  • Plan mode for multi-file changes — Shift+Tab to toggle
  • Subagents for heavy lifting — delegate exploration, tests, docs

Before Ending

  • Run /wrap-up — audit changes, capture learnings
  • Or at minimum: git status, check for uncommitted work
  • Generate /handoff if continuing later

Every Day

Update

  • Keep Claude Code updated (it auto-updates, but check)
  • Read the changelog for new features
  • Review your LEARNED.md — prune stale rules

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

  • /usage to check plan limits
  • /cost for API key users
  • Watch for fast mode billing (charges extra usage from first token)

Every Week

Review Patterns

  • Run /insights to see correction trends
  • Identify hot learnings (corrected often, not learned)
  • Prune cold learnings (learned but never applied)

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

  1. Try a different model — switch to Opus for hard problems
  2. Use ultrathink in your prompt for maximum reasoning
  3. Provide a screenshot (Claude is multimodal)
  4. Share browser console logs via MCP
  5. Start a fresh session with more targeted context

When Context Is Degraded

  1. Run /compact manually
  2. Or start a new session with /resume
  3. Set CLAUDE_AUTOCOMPACT_PCT_OVERRIDE=50 for proactive compaction
  4. Use subagents to isolate heavy output (test results, logs)

When You're Blocked

  1. /doctor for configuration issues
  2. Try a different approach — don't brute force
  3. Use the debugger agent for systematic investigation
  4. Check if it's a model limitation vs. a prompt issue
  5. Ask Claude to "think step by step about why this isn't working"

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:

  • /voice to enable
  • Hold spacebar → speak → release
  • Great for: describing bugs, architecture discussions, code reviews
  • Mix with typed input for precise file paths

Tool Layering

The most productive setup uses multiple tools:

Primary editor (Cursor/VS Code)
  ├── Tab completions for small edits
  ├── Inline chat for quick questions
  └── Claude Code in terminal for:
      ├── Multi-file changes
      ├── Debugging hard issues
      ├── Architecture decisions
      └── CI/CD and git operations

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