tool

Claude Code — Coding Agent Cheatsheet

Practical setup and operating guide for Claude Code: install the CLI, ground it in a repository, wire memory, hooks, MCP tools, and safe review loops.

2026-09-25 (3 weeks ago)
5 steps
~55 min total
Claude Code current

claude-code
anthropic
coding-agent
cli
hooks
memory
mcp

Claude Code — Coding Agent Cheatsheet

One-line: Claude Code is Anthropic’s repo-grounded coding-agent CLI for reading, editing, testing, and reviewing software with persistent project context, hooks, MCP tools, and explicit permissions.

Installation

Use the official Claude Code documentation for the latest install path, then verify the CLI before letting it modify a repository.

# Install or update from the official Claude Code distribution.
claude install stable

# Start Claude Code in the repository you want it to understand.
cd path/to/project
claude

First-run checklist:

Authenticate with an Anthropic account or configured provider.
Open the project root, not a random subdirectory.
Run /doctor if login, shell, hooks, MCP, or settings do not load.
Run /context to inspect loaded files, memory, skills, and MCP tools.

If claude install is not available in your environment, follow the current install page linked in Reference and keep this cheatsheet as the operating workflow.

Step-by-Step Setup & Optimization

Step 1 — Ground Claude in the repository

Goal: A CLAUDE.md at the repo root carrying only durable project factsTime: ~10 minLevel: beginner

Create or update CLAUDE.md at the repo root with only durable facts:

# Project guide
- Package manager: pnpm
- Checks before PR: pnpm test && pnpm type-check
- Do not edit generated files under dist/
- Prefer small commits and include the issue number in commit messages

Use /memory and /context to confirm what loaded. Keep memory short; project rules belong in CLAUDE.md, while personal preferences belong in user memory.

/memory and /context show your CLAUDE.md content loaded.

Step 2 — Choose a permission posture

Goal: A permission level matched to how much you trust the repoTime: ~5 minLevel: beginner

Start strict, then relax only for trusted repos.

Situation	Recommended posture
Unknown repo	Read-only exploration, approve every shell command
Normal feature work	Workspace writes plus explicit command approvals
CI-style automation	Allowlisted commands and hooks that fail unsafe actions
Secrets or production	Separate sandbox, no broad filesystem or network access

Your posture matches the table — strict on unknown repos, relaxed only where trusted.

Step 3 — Add deterministic hooks

Goal: Guardrails enforced by hooks instead of prompt wordingTime: ~15 minLevel: intermediate

Hooks make repeated guardrails deterministic instead of prompt-dependent. Common hooks run formatters after edits, reject dangerous commands, or notify when a task finishes.

{
  "hooks": {
    "PostToolUse": [
      {
        "matcher": "Edit|Write",
        "hooks": [{ "type": "command", "command": "pnpm prettier --write ${file}" }]
      }
    ]
  }
}

Keep hook commands fast and idempotent. If a hook can delete files, publish packages, deploy, or rewrite history, require human review.

Make an edit and watch the hook fire — e.g. the formatter runs on the saved file.

Step 4 — Connect external context with MCP

Goal: External tools exposed to Claude through typed MCP interfacesTime: ~15 minLevel: intermediate

Use MCP servers for tools Claude should discover and call through a typed interface: issue trackers, docs search, databases, observability, or local project utilities.

/mcp          # inspect connected servers and tools
/doctor       # debug MCP, settings, hooks, and memory loading

Prefer narrow MCP servers over broad shell access. Name tools by user intent, document inputs, and log side effects.

/mcp lists your connected servers and tools; /doctor reports no loading errors.

Step 5 — Work in small issue loops

Goal: A repeatable loop that keeps every change reviewableTime: ~10 minLevel: intermediate

Ask Claude to restate the issue and acceptance criteria.
Request a short plan before edits.
Let it edit one coherent slice.
Run the project checks.
Review the diff yourself before commit or PR.

Each loop ends with project checks passing and a diff you reviewed yourself.

Best Practices

Do

✅ Keep CLAUDE.md concise, versioned, and specific to the repository.
✅ Ask for a plan, then make Claude execute one small step at a time.
✅ Run /context when behavior looks surprising; loaded context often explains it.
✅ Use hooks for formatting, policy checks, and notifications.
✅ Use MCP for scoped integrations instead of giving blanket shell/network freedom.
✅ Commit after known-good milestones so agent mistakes are easy to revert.

Don’t

❌ Paste secrets into chat or memory files.
❌ Put temporary issue details into durable project memory.
❌ Let hooks run destructive commands without review.
❌ Start huge refactors without tests, a rollback point, and a review plan.
❌ Assume Claude saw every file; explicitly attach or reference critical files.

Quick Command Reference

Command	Use
`claude`	Start Claude Code in the current project.
`claude install stable`	Install or reinstall the stable CLI when supported.
`/context`	Inspect loaded system, memory, skills, MCP, and conversation context.
`/doctor`	Diagnose settings, hooks, MCP, login, and project configuration.
`/memory`	View or update memory deliberately.
`/mcp`	Inspect configured MCP servers and exposed tools.
`/hooks`	Inspect or debug configured lifecycle hooks.

Expected Outcomes

After setup, you should have:

A repository-level CLAUDE.md that captures durable project rules.
A safe default permission model for reads, edits, shell commands, and network use.
Hooks that handle repeatable quality gates without relying on the model to remember.
MCP integrations that expose useful external context through narrow interfaces.
A repeatable issue workflow: plan, edit, test, review, commit, and PR.

Reference

Sources & deeper reading