TechRounder Developer Field Guide

Claude Code, decoded for the moment you need it

Find the right Claude Code command for setup, daily work, automation, security, extensions, and recovery—along with clear examples and cautions that help you use it correctly.

Find a command Follow a workflow

Updated July 28, 2026 · Fast-changing commands are marked “Verify live”.

five-command-survival-kit

Tap any command to copy it.

165commands and shortcuts

13practical categories

34MCP recipes

Copy readyexamples and workflows

01 · Route by outcome

What are you trying to do?

These routes jump directly into a filtered command set, so you can work from intent instead of memorizing syntax.

02 · Universal lookup

Command Explorer

Search commands, descriptions, aliases, examples, and keywords. Press / anywhere on this page to focus search.

Search the complete reference

/

Category Type Safety Freshness

03 · Compose, don’t memorize

Headless command builder

Build a bounded automation command and see exactly what every added flag changes.

Task prompt

Output Maximum turns Model (optional)

Pre-approved tools Read Grep Glob Bash

Generated command

Safety note: pre-approval removes prompts for the named tools. Grant only the minimum needed by the automation.

04 · Mental model

Configuration precedence

Higher layers win. Use the lowest layer that matches the intended audience.

Enterprise/etc/claude-code/managed-settings.json

Highest priority; enforced by administrators.

Project local.claude/settings.local.json

Personal to this project; normally Git-ignored.

Project shared.claude/settings.json

Team configuration committed with the project.

User global~/.claude/settings.json

Personal defaults across projects.

05 · Persistent instructions

Memory map

CLAUDE.md is a hierarchy, not one magic file. Keep shared rules concise and place scope-specific detail close to the code it governs.

Organization policy/etc/claude-code/CLAUDE.md

Managed instructions for the whole organization.

User global~/.claude/CLAUDE.md

Personal instructions used across projects.

Project shared./CLAUDE.md or ./.claude/CLAUDE.md

Team rules that belong in version control.

Project personal./CLAUDE.local.md

Private project preferences; Git-ignored.

Path-scoped rules.claude/rules/*.md

Rules with paths frontmatter that load only for relevant files.

Auto-memory~/.claude/projects/.../MEMORY.md

Persistent notes; the first 200 lines load automatically in supported versions.

06 · Working recipes

Four dependable workflows

Select a playbook to reveal a compact sequence. Every command in a step can be copied.

Start a project safely

Understand the repo before allowing edits.

01
Launch inside the project root.
02
Create a concise project briefing if one does not exist.
03
Move into Plan Mode for the first architectural pass.
04
Review tool boundaries before implementation.

07 · Choose intelligence

Model aliases

Aliases are intentionally versionless. Use /model to see what they resolve to in your account.

fable

Claude Fable 5

Flagship / hardest architectural tasks

Availability can vary by account and release; check /model.

opus

Claude Opus

Complex reasoning and architecture decisions

The exact version behind this alias can change.

sonnet

Claude Sonnet

Balanced everyday coding

The exact version behind this alias can change.

haiku

Claude Haiku

Fast, inexpensive, tightly scoped work

The exact version behind this alias can change.

best

Automatic selection

Strongest available option

Check /model to confirm availability.

opusplan

Opus → Sonnet

Opus planning followed by Sonnet execution

Check /model to confirm availability.

lowFast, latency-sensitive tasks.

mediumCost-sensitive work with moderate reasoning.

highBalanced default for most demanding work.

xhighDeeper reasoning with higher token use.

maxDeepest session-only reasoning; may overthink simple work.

ultracodeNewer orchestration-focused mode; verify availability before use.

08 · Keyboard layer

Shortcut wall

Terminal and IDE mappings vary. Press ? in a live session for the local truth.

Shift + TabCycle permission modes: Default → Accept Edits → Plan (and any additional enabled modes).

!Enter Bash mode for direct shell commands.

@Mention a file or folder in the prompt.

/Open the slash command menu.

\ + EnterInsert a line break without submitting.

EscInterrupt Claude mid-response.

Esc, EscOpen the rewind/checkpoint menu.

Ctrl + CClear input; press again on an empty prompt to exit.

Ctrl + DExit the session.

Ctrl + LClear and redraw the terminal.

Ctrl + OToggle the verbose transcript/thinking view.

Ctrl + RReverse-search prompt history.

Ctrl + TToggle the background task list.

Ctrl + BBackground a running Bash command.

Ctrl + VPaste an image from the clipboard.

Ctrl + GOpen the prompt in your $EDITOR.

Ctrl + X, Ctrl + KKill all background agents after double confirmation.

Option + PSwitch the active model.

Option + TToggle extended thinking.

Option + OToggle fast mode.

Cmd + EscQuick-launch Claude in supported IDEs.

↑ / ↓Move through prompt history.

?Show the shortcut list for the current terminal or IDE.

09 · Vocabulary

Glossary without the jargon spiral

Short definitions for the concepts that make the rest of the reference easier to use.

Session

One continuous run of claude in a project. It has its own chat history.

Context window

The model’s working memory: prompts, replies, files, and tool results currently in scope.

Token

The unit used to measure model input and output; roughly three-quarters of an English word.

CLAUDE.md

An automatically loaded Markdown briefing for project conventions, commands, and constraints.

Plan Mode

A look-before-touching mode used to explore and propose without editing or executing.

Accept Edits Mode

A mode that auto-approves file edits while other risky actions may still need confirmation.

Permissions

Rules that decide what Claude may do automatically, what requires approval, and what is denied.

Tool

A capability such as reading a file, editing, searching, or running Bash.

MCP

Model Context Protocol: a standard for exposing external systems to Claude as tools.

Subagent

A focused Claude instance with separate context that reports back to the main session.

Hook

A command triggered at a lifecycle event such as before or after a tool call.

Skill

A reusable workflow package led by SKILL.md and invoked when its description matches the task.

Command

An instruction beginning with /; built-in commands ship with Claude Code and custom workflows can add more.

Plugin

A shareable bundle that can include commands, agents, skills, hooks, and MCP servers.

Worktree

A second working copy of a Git repository on another branch, useful for isolated parallel work.

10 · Setup

Install and launch

Pick the method that matches the machine, then run claude doctor if authentication or environment checks fail.

macOS / Linux · recommended

macOS · Homebrew

Windows · PowerShell

Node.js 18+

11 · Extension anatomy

Skills, subagents, custom commands, and hooks

These mechanisms overlap, but they solve different problems. Use the smallest one that expresses the behavior clearly.

Custom slash commandYou invoke it

Store a prompt in .claude/commands/name.md or ~/.claude/commands/name.md. Use $ARGUMENTS, $1/$2, @filename, and pre-prompt shell commands where supported.

Address this bug: $ARGUMENTS.
Refer to @CLAUDE.md for project rules.

SkillClaude invokes it

Put SKILL.md and optional helpers under .claude/skills/<name>/ or ~/.claude/skills/<name>/. Discovery is driven by the skill description.

---
name: generating-commit-messages
description: Generate clear commit messages from staged diffs.
---

1. Run git diff --staged
2. Explain what and why

SubagentFocused delegated context

Define Markdown with YAML frontmatter under .claude/agents/ or ~/.claude/agents/. Restrict tools to the minimum required.

---
name: code-reviewer
description: Review changed code for security and maintainability.
tools: Read, Grep, Bash
model: inherit
---

Lifecycle hookDeterministic shell automation

Hooks fire on events and run with your credentials. Review every script. Exit code 0 reports success; exit code 2 blocks the action and returns feedback to Claude.

PreToolUse PostToolUse UserPromptSubmit Notification Stop SubagentStop PreCompact SessionStart SessionEnd Post-session

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

12 · Guardrails

Default to narrow permissions

Permission rules generally fall into allow, ask, and deny. Keep sensitive reads denied, require confirmation for network or publishing actions, and pre-approve only routine commands you understand.

MCP wildcards such as mcp__github__* are not valid; approve a server or an exact tool name instead.
Approve the server with mcp__github or one tool with mcp__github__get_issue.
Enterprise deniedMcpServers rules take precedence over allow rules.
Hooks execute as you. A malicious hook can damage files or expose data.

Conservative project starter

{
  "permissions": {
    "allow": [
      "Bash(git status)",
      "Bash(git diff)"
    ],
    "ask": [
      "Bash(git push:*)",
      "Bash(npm install *)"
    ],
    "deny": [
      "Read(./.env*)",
      "Read(./secrets/**)",
      "Read(./**/credentials*)",
      "Bash(rm -rf:*)",
      "Bash(curl:*)"
    ]
  }
}

Claude Code, decoded for the moment you need it

What are you trying to do?

Command Explorer

Headless command builder

Configuration precedence

Memory map

Four dependable workflows

Start a project safely

Recover from a wrong turn

Automate a review

Run parallel feature work

Model aliases

Shortcut wall

Glossary without the jargon spiral

Session

Context window

Token

CLAUDE.md

Plan Mode

Accept Edits Mode

Permissions

Tool

MCP

Subagent

Hook

Skill

Command

Plugin

Worktree

Install and launch

Skills, subagents, custom commands, and hooks

Default to narrow permissions