Claude Code cheatsheet

The commands and patterns I actually use day to day. Not exhaustive -- just what earns its keep.

Slash commands

Command	What it does
/clear	Clear the conversation. Start fresh without restarting the CLI.
/init	Generate a CLAUDE.md for the current repo by scanning the codebase.
/help	Show available commands.
/memory	View and manage persistent memory.
/config	Open settings (theme, model, etc.).
/loop <interval> <prompt>	Run a prompt on a recurring interval (e.g. /loop 5m /check-deploys).
/run	Launch the project and verify a change works in the real app.
/verify	Run the code and observe behavior to confirm a fix.
/security-review	Security review of pending changes on the current branch.
/review	Review a pull request.

File references

• @<path> -- reference a file in your message (e.g. @src/app.ts please explain). Claude reads it directly instead of you pasting.
• @<folder>/ -- reference a folder; useful when you want Claude to scan a small directory.

Project configuration

File	Purpose
CLAUDE.md	Project instructions checked into the repo. Loaded into every session.
CLAUDE.local.md	Local-only instructions, not checked in. Good for personal preferences and secrets-free overrides.
.claude/settings.json	Permissions, environment variables, hooks. Checked in.
.claude/settings.local.json	Same as above but local-only, gitignored.
.claude/skills/<name>/SKILL.md	A reusable workflow Claude can invoke. Built up as recurring patterns emerge.
.claude/rules/*.md	Auto-loaded instruction files (e.g. communication style, testing rules).

Permissions cheatsheet

In .claude/settings.json:

{
  "permissions": {
    "allow": [
      "Bash(npm run *)",
      "Bash(git status)",
      "Bash(git diff*)",
      "Edit(src/**)",
      "Write(docs/**)"
    ],
    "deny": [
      "Bash(rm -rf *)",
      "Bash(git push --force*)"
    ]
  }
}

Patterns use glob-style wildcards. Bash(npm run *) allows any npm run subcommand without prompting.

Hooks (automated behaviors)

Settings.json hooks run shell commands on events (Stop, PreToolUse, PostToolUse, etc.). Use them when you need a behavior to fire deterministically -- memory and CLAUDE.md instructions are advisory, hooks are not.

{
  "hooks": {
    "Stop": [
      {"matcher": "", "hooks": [{"type": "command", "command": "notify-send 'Claude done'"}]}
    ]
  }
}

Plan mode

Useful when the change is non-trivial. Claude explores, writes a plan to a file, asks clarifying questions, and waits for approval before touching code.

• Toggle: shift+tab (or whatever your keybind is set to)
• Plan file lives at ~/.claude/plans/<slug>.md
• Exit plan mode = approval to start coding

Subagents

Spawn specialized agents for parallel or isolated work:

• Explore -- fast read-only search, returns excerpts. Good for "where is X defined."
• Plan -- software-architect-style implementation planning.
• general-purpose -- broad multi-step work.

Custom agents go in .claude/agents/<name>.md with a description field. Trigger by description match or via Agent(subagent_type: "<name>").

Memory

Memory persists across conversations and lives outside the project:

~/.claude/projects/<encoded-cwd>/memory/
  MEMORY.md          ← index, always loaded
  user_role.md       ← who you are
  feedback_*.md      ← how you like to work
  project_*.md       ← project context

Trigger phrases:

• "remember that I always want X" -- saves a feedback memory
• "forget that" -- finds and removes the relevant entry

Useful Bash patterns

# Resume a previous session
claude --continue

# Run a one-off prompt non-interactively
claude --print "show git status and summarize"

# Use a specific model
claude --model claude-sonnet-4-6

Tips that compound

• Start with /init on a new repo. Generated CLAUDE.md is a solid baseline.
• Use plan mode for anything non-trivial. Cheap insurance.
• Build skills the second time you do something, not the first. Speculative skills go stale.
• Permissions allowlist > confirming every command. Lower friction = you actually use the tool.
• Trust but verify subagent output. Their summaries describe intent, not results.