Cheat Sheets - Claude Code

Fire and forget

Small, unambiguous tasks. Just describe what you want and let it go.

Plan Mode

Medium features where you want alignment before Claude touches code. Review the plan, edit it, then execute.

Interview mode

Fuzzy requirements. Ask Claude to interview you with AskUserQuestion before writing a spec or starting work.

Two-Claude review

One Claude drafts the plan, a second reviews it "as a staff engineer." Catches blind spots before they become bad code.

Subagents

Offload isolated, context-heavy subtasks (security review, test writing, research) to keep your main context clean.

Parallel worktrees

Independent workstreams running simultaneously in separate git worktrees. Each Claude session stays isolated.

"Ralph Wiggum"

Large, well-specced body of work you want to run unattended. Requires sandboxing, solid specs, and working backpressure (tests).

This approach leverages the `--dangerously-skip-permissions` flag that bypasses all permission prompts for these sessions, so only use this in a sandboxed environment (Docker container with minimal credentials, no access to SSH keys/browser cookies/production secrets).

Slash commands

Anything you do repeatedly in a day. Saved in .claude/commands/, checked into git.

CLAUDE.md

Your persistent project brain. Every time Claude does something wrong, add a guardrail so it doesn't repeat.

GitHub Actions integration

Trigger Claude Code from Slack, Jira, or CI events to produce a tested PR with no manual session management.