Claude Code usage tips and tricks

calude.md

• Write ultra-precise prompts — state edge cases, implicit assumptions, and concrete examples; Claude can’t guess them.
• Ask Claude to “ultrathink” when deeper multi-step reasoning is needed.
• Use planning mode (Shift+Tab twice) or another model for high-level design before coding.
• Delegate with the Task tool — spawn specialist sub-agents (design, research, etc.) to tackle parts in parallel.
• Make Claude RTFM first — link docs or start a research Task so it learns the latest best practices before scaffolding.
• Start new threads often — run /clear to reset context and cut drift when you switch topics.
• Restart with --resume to branch without losing work.
• Use claude --dangerously-skip-permissions" to avoid constant prompts, but 1) it's probably better run in a container 2) mind the limits, it will eat them fast
• Dictate long prompts with voice-to-text; Claude copes well with messy transcripts.
• Stage early, stage often — git add after each good turn to simplify rollbacks.
• Leverage git worktrees to run multiple Claude sessions on one repo without clobbering each other.
• Iterate instead of one-shotting big refactors—small feedback loops beat giant prompts that compound errors.
• Interrupt fast — tap Esc to stop runaway actions and course-correct immediately.
• Create custom slash commands in .claude/commands for prompts you reuse.
• Paste images directly to get quick visual feedback on UI changes or install MCP that takes screenshots for you(but this wastes tokens)