This post is about using the actual claude CLI tool itself, not about using the best prompt techniques or plugins or whatever. I'll maybe do a plugins post later.

Table of contents

• Terminology
• Launch headless Claude, ex in Kubernetes cron jobs
• Skills & Agents
  • When to use Skills vs Agents
  • Built-in agents/subagents
  • Built-in skills
  • Defining Claude Agents
  • Creating custom Skills
    • Where they live
    • Frontmatter
    • Placeholders
    • Shell injection
  • Managing Permissions
• Slash commands reference
  • Integrations and misc
• Keyboard shortcuts
  • General
  • Input and editing
  • Multiline input
  • Prefix triggers
• Worktrees
• Session Management
  • Forking sessions
  • When to set Session ID manually
  • Session management and viewing
• Hooks
• CLAUDE.md and Rules
  • CLAUDE.md file discovery
  • Rules - Conditionally applied CLAUDE.md

Terminology

Useful to know what these are

Term	What it means
Turn	One back-and-forth with the model. Claude generates a response (possibly with tool calls), reads the results, then generates again. Each of those generation steps counts as a turn.
Context (or context window)	The running transcript the model can see on a given turn: your messages, its messages, tool calls, tool results, system prompt, CLAUDE.md, loaded skills. Measured in tokens.
Compaction	When the context window fills up, Claude Code summarizes older history in-place to free space. Triggered automatically or via /compact.
Headless / print mode	Running claude -p "prompt" so it answers and exits instead of opening the TUI. The entry point for scripts and CI.
Frontmatter	The YAML block between --- markers at the top of a markdown file. Carries config like name, description, tools.
Hook	A shell command Claude Code runs at a lifecycle event (tool use, session start, etc.) to allow, block, or modify the action. Deterministic, not LLM-based.
Subagent	A child agent spawned from the main conversation. Runs in its own context window with its own tools and system prompt. Returns a summary.
Plan mode	Read-only exploration mode. Claude can search and read but cannot edit or run shell commands.

Launch headless Claude, ex in Kubernetes cron jobs

Install it in a Dockerfile. I like to launch it from k8s in a CronJob from there.

vi Dockerfile

FROM node:22-slim
RUN npm i -g @anthropic-ai/claude-code

Collect the value for CLAUDE_CODE_OAUTH_TOKEN.

claude setup-token

Run the container in k8s with CLAUDE_CODE_OAUTH_TOKEN env var set. Execute the claude command in prompt mode.

claude \
  --output-format json \
  --m opus \
  -p $CLAUDE_PROMPT

You'll probably also want to use --permission-mode auto or --dangerously-skip-permissions too so it doesn't freeze on permission requests. Consider some of these launch flags:

Flag	What it does
-p, --print	Non-interactive mode: print the response and exit. The SDK / CI entry point.
--output-format <text|json|stream-json>	Print-mode output format.
--input-format <text|stream-json>	Print-mode input format.
-m, --model <alias|id>	Pick model for this session (sonnet, opus, haiku, or a full ID like claude-opus-4-7).
--effort <level>	low, medium, high, xhigh, max. Works on Opus 4.7, Opus 4.6, Sonnet 4.6.
--mcp-config <path|json>	Load MCP servers from file(s) or a JSON string. Repeatable.
--strict-mcp-config	Only use servers from --mcp-config, ignore project/user configs.
--json-schema '<schema>'	Constrain print-mode output to match a JSON Schema.
--max-turns <n>	Cap agentic turns in print mode. Exits with error when hit.
--max-budget-usd <amount>	Cap API spend in print mode.
--no-session-persistence	Don't write session to disk. Print mode only.
--system-prompt "text"	Replace the default system prompt.
--append-system-prompt "text"	Append to the default system prompt.
--system-prompt-file <path>	Same, from a file. Paired --append-system-prompt-file exists.
--verbose	Full turn-by-turn output.
--debug [categories]	Debug mode, optional filter like "api,hooks" or "!statsig".

Note: --system-prompt and --system-prompt-file are mutually exclusive; --append-* variants compose with either.

Then run it. Another example:

claude \
  -p "$CLAUDE_PROMPT" \
  --model sonnet \
  --output-format json \
  --permission-mode bypassPermissions \
  --max-turns 10 \
  --max-budget-usd 2.00 \
  --no-session-persistence

This'd give you JSON output for parsing (datadog etc), a hard cap on agentic turns, a dollar ceiling so a runaway job can't drain your account, and no session file written to disk. Drop the --max-* flags if you trust the prompt, swap bypassPermissions for auto if you want the auto-mode classifier to gate risky calls.

---

Skills & Agents

When to use Skills vs Agents

Skills are procedures, agents are workers.

A skill is a directory containing a SKILL.md prompt, plus any scripts, templates, or reference docs it ships alongside. When the skill triggers, only SKILL.md gets injected into your current conversation as an extra message. Claude sees it alongside everything else already in the chat, and the skill's instructions apply for the rest of the session. The supporting files sit in the skill directory and don't enter context unless SKILL.md tells Claude to read them with the Read tool or run them via Bash.

An agent runs in its own context window with its own system prompt, tool allowlist, model, and permissions. It works independently and returns only a summary. Only the agent's name and description sit in the parent conversation's context. The body stays out until the agent actually runs, and even then it only shows up in the agent's own window.

Rule of thumb:

If the work...	Lean
Is a repeatable recipe (commit style, how to run a migration)	Skill
Produces a lot of intermediate output you won't reference again	Agent
Needs to read what's already in the conversation	Skill
Should run on a different model (Haiku for cost, Opus for brains)	Agent
Needs a tighter tool or permission scope than the parent	Agent
Bundles shell scripts, templates, or reference files	Skill
Iterates autonomously over many turns before returning	Agent
Is one short instruction block that could live in CLAUDE.md	Skill

Default to a skill. If it keeps dumping logs or file contents into your main window, promote it to an agent. The cost of a skill that should have been an agent is a bloated parent context. The cost of an agent that should have been a skill is an extra round trip and less visibility into what actually happened.

Built-in agents/subagents

I use the term "agent" and "subagent" pretty interchangeably with Claude, and I think I'm doing that right. If there's one different, a subagent is called within a conversation where an agent is called as a launch arg with --agent.

Name	Model	Tools	Purpose
Explore	Haiku	read-only	File discovery, code search
Plan	inherit	read-only	Codebase research during plan mode
general-purpose	inherit	all	Multi-step tasks
statusline-setup	Sonnet	n/a	Used by /statusline
Claude Code Guide	Haiku	n/a	Answers feature questions

Built-in skills

Command	What it does
/simplify [focus]	Review changed files and fix quality/efficiency issues.
/batch <instruction>	Orchestrate large-scale parallel changes across the codebase.
/debug [description]	Enable debug logging and troubleshoot.
/loop [interval] [prompt] (alias /proactive)	Run a prompt repeatedly.
/claude-api	Load Claude API reference for the session's language.
/fewer-permission-prompts	Scan transcripts and propose allowlist entries.

Defining Claude Agents

Agents, in this context, are just .md files with a special spec. You can find them in two places:

• Project (repo): .claude/agents/<name>.md
• Personal (global): ~/.claude/agents/<name>.md

An agent file is just some YAML frontmatter metadata props and a markdown body containing the prompt. Only name and description are required.

Parent conversations load each available subagent's name and description so the orchestrator can decide when to delegate.

---
# Required
name: code-reviewer                 # lowercase + hyphens, unique
description: Reviews diffs for bugs, style, and security. Use after any code change.

# Tool and capability gates
tools: [Read, Grep, Bash]           # omit to inherit everything
disallowedTools: [Bash(rm *)]       # subtractive, applied after `tools`
mcpServers: [slack]                 # server name, or inline MCP config
skills: [code-review]               # full skill content gets injected at startup

# Model and execution
model: sonnet                       # sonnet|opus|haiku|<full-id>|inherit (default: inherit)
effort: medium                      # low|medium|high|xhigh|max
maxTurns: 20                        # agentic turn cap
background: false                   # true = always run as a background task
isolation: worktree                 # runs in a throwaway git worktree

# Permissions and lifecycle
permissionMode: acceptEdits         # default|acceptEdits|auto|dontAsk|bypassPermissions|plan
hooks:                              # PreToolUse | PostToolUse | Stop, scoped to this agent
  PreToolUse:
    - matcher: Bash
      command: ./scripts/log-tool-use.sh
memory: project                     # user|project|local persistent memory

# UX
color: blue                         # red|blue|green|yellow|purple|orange|pink|cyan
initialPrompt: "Start by listing changed files."  # auto-submitted when run via --agent
---

You are a code reviewer. Read the diff, then...

Scope precedence, highest to lowest: managed org settings, --agents CLI flag, .claude/agents/ (project), ~/.claude/agents/ (personal), plugin agents.

You can call agents from within claude conversationally, or from the command-line as entrypoints with claude --agent <name>.

Creating custom Skills

Skills are the native Claude Code way to package reusable workflows. They replaced .claude/commands/*.md (legacy commands still work but skills take precedence on name collision).

Where they live

Location	Scope
Managed settings	All users in the org
~/.claude/skills/<name>/SKILL.md	Personal
.claude/skills/<name>/SKILL.md	Project
Plugin skills/<name>/SKILL.md	Where the plugin is enabled

Frontmatter

Field	Notes
name	Slash command name. Defaults to the directory name.
description	Claude reads this to auto-invoke. Recommended.
when_to_use	Extra trigger context; appended to description.
argument-hint	Shown in autocomplete (e.g. [issue-number]).
disable-model-invocation	true = user-only; Claude won't auto-load it.
user-invocable	false = hide from the / menu.
allowed-tools	Tools auto-approved while the skill is active.
model, effort	Overrides for this skill.
context	fork to run in an isolated subagent.
agent	Which subagent when context: fork.
hooks	Hooks scoped to this skill.
paths	Globs limiting when the skill auto-activates.
shell	bash (default) or powershell for !command blocks.

Placeholders

• $ARGUMENTS: all args passed at invocation.
• $ARGUMENTS[N] or $N: specific argument, 0-indexed.
• ${CLAUDE_SESSION_ID}, ${CLAUDE_SKILL_DIR}.

Shell injection

!`command` inside skill content runs the command before the skill is sent to Claude; the output replaces the placeholder. Multi-line uses a fenced ```! block. Turn off globally with disableSkillShellExecution: true in settings.

That has the security implications you'd expect.

Managing Permissions

Permission prompts can slow you down. For a while I tried to grant permissions to every litle thing, feels best-practice, but its so slow. I mostly just use permission-mode auto now.

Flag	What it does
--permission-mode <mode>	default, acceptEdits, plan, auto, dontAsk, bypassPermissions.
--dangerously-skip-permissions	Equivalent to starting in bypassPermissions. Read the name.
--allowedTools "Bash(git log *)" Read	Tools that run without prompting. Space-separated list.
--disallowedTools "Bash(curl *)"	Tools removed from context entirely. Applied before allow.
--tools "Bash,Edit,Read"	Restrict the built-in tools Claude can see. "" disables all of them.
--add-dir <path>	Grant file access to an additional directory outside CWD.

---

Slash commands reference

There are a lot of these, here's some good ones.

• /btw is great, use it when you think claude is stuck
• I pretty much only use xhigh effort currently
• I personally use /coderabbit for PR reviews but /review is nice too. It figures out your current PR itself.

Command	What it does
/clear (aliases /reset, /new)	Start a fresh conversation. Previous one is still resumable.
/compact [instructions]	Summarize to free context. Optional hint focuses the summary.
/context	Visualize current context usage as a colored grid.
/model [alias]	Change model. Left/right arrows adjust effort in the picker.
/effort [level]	low/medium/high/xhigh/max/auto. No arg opens a slider.
/plan [description]	Enter plan mode.
/agents	Open subagent manager. Left and right to nav. CRUD or run from here.
/hooks	View hook configs.
/permissions (alias /allowed-tools)	Manage allow/ask/deny rules.
/memory	Edit CLAUDE.md files, toggle auto-memory, view auto-memory entries.
/mcp	Manage MCP server connections and OAuth.
/skills	List available skills. Press t to sort by token count.
/resume (alias /continue)	Resume a conversation by ID/name or via picker.
/rewind (aliases /checkpoint, /undo)	Rewind conversation and code to a previous point.
/branch [name] (alias /fork)	Branch current conversation.
/btw <question>	Ask a side question that doesn't get added to history.
/review [PR]	Review a PR locally.
/security-review	Analyze pending branch changes for security issues.
/cost	Token usage stats.
/usage	Plan usage limits and rate-limit status.
/doctor	Diagnose the installation.

Integrations and misc

Some notes from my testing of these:

• /voice is novel but way worse than ChatGPT at recognizing what you said. I don't find it very usable.
• /install-github-app is neat in that you don't need other infra. See instructions.

Command	What it does
/ide	Manage IDE integrations.
/chrome	Configure Claude in Chrome.
/desktop (alias /app)	Continue in Claude Code Desktop.
/mobile (aliases /ios, /android)	QR code to download the mobile app.
/install-github-app	Set up Claude GitHub Actions app.
/install-slack-app	Install Claude Slack app.
/remote-control (alias /rc)	Expose session to remote control.
/teleport (alias /tp)	Pull a web session into this terminal.
/ultraplan <prompt>	Draft plan in ultraplan session, execute remotely.
/ultrareview [PR]	Multi-agent cloud code review.
/autofix-pr [prompt]	Spawn web session to push fixes when CI fails.
/schedule [description] (alias /routines)	Create/list/run routines.
/plugin, /reload-plugins	Plugin management.
/sandbox	Toggle sandbox mode (supported platforms only).
/voice	Push-to-talk voice dictation.

---

Keyboard shortcuts

General

Key	Effect
Esc	Cancel the current generation.
Esc Esc	Rewind to a previous checkpoint (conversation and code).
Shift+Tab (or Alt+M)	Cycle permission modes: default → acceptEdits → plan → auto → bypassPermissions.
Ctrl+C	Cancel current input or generation.
Ctrl+D	Exit the session.
Ctrl+L	Clear the prompt input and redraw (keeps history).
Ctrl+R	Reverse search command history.
Ctrl+O	Toggle transcript viewer.
Ctrl+B	Background the running bash task. In tmux, press twice.
Ctrl+T	Toggle task list in the status area.
Ctrl+X Ctrl+K	Kill all background agents. Press twice within 3s to confirm.
Ctrl+G or Ctrl+X Ctrl+E	Open the current prompt in $EDITOR.
Option+P / Alt+P	Switch model without clearing the prompt.
Option+T / Alt+T	Toggle extended thinking.
Option+O / Alt+O	Toggle fast mode.

Input and editing

Key	Effect
Ctrl+A / Ctrl+E	Start / end of line.
Ctrl+K	Delete to end of line.
Ctrl+U	Delete to start of line.
Ctrl+W	Delete previous word.
Ctrl+Y	Paste last deleted text.
Alt+Y (after Ctrl+Y)	Cycle paste history. Requires "Use Option as Meta Key" in iTerm
Alt+B / Alt+F	Back / forward a word. Requires "Use Option as Meta Key" in iTerm

Multiline input

Terminal	Newline binding
Any terminal	\ then Enter
macOS default	Option+Enter
iTerm2, WezTerm, Ghostty, Kitty	Shift+Enter (no config needed)
VS Code, Alacritty, Zed, Warp	Run /terminal-setup to install Shift+Enter
Control sequence (anywhere)	Ctrl+J

Prefix triggers

Prefix	What it does
/	Slash commands and skills.
!	Bash mode: run a shell command and add the output to context.
@	File path autocomplete.
Hold Space	Push-to-talk dictation (if /voice is on).

---

Worktrees

A worktree is git concept, it basically lets you work on more than one branch at a time. You'll see it saved to .claude/worktrees/. Don't forget to gitignore them.

Flag	What it does
-w, --worktree <name>	Start in a git worktree at <repo>/.claude/worktrees/<name>.
--tmux	Pair with --worktree to open in a tmux pane.

I personally use this with multiple tmux sessions across a single repo so I can work on more than one thing at a time. You do end up with the usual merge conflict situation, but claude's pretty good about sorting that out on its own too.

I'm personally a tmux nerd, I'm genuinely just a happier person with tmux running, so the --tmux flag existing is exciting to me. It's kind of janky though, it claims it should open a pane but at the time of my writing this it opens it into a new session and puts the original pane you ran it in into a blocking state, so personally I'd just manage the worktrees myself for now.

---

Session Management

Often you need to quit to reload MCP tools or reboot. You might also want to just revisit an prior conversation. Sessions are saves of where your context left off.

Show current session UUID:

ls -t ~/.claude/projects/$(pwd | sed 's|/|-|g')/*.jsonl | head -1

When you exit a claude session it'll often print the --resume command for you at the bottom. If you missed it, you can use claude --continue to jump back in:

claude
> do the thing
> exit
claude --continue

Flag	What it does
-c, --continue	Resume the most recently exited conversation in the current directory. Basically -r with the arg pre-filled
-r, --resume [name|id]	Resume a specific session by name or UUID, or open an interactive picker.
--fork-session	With --resume or -c, create a new session ID instead of reusing the old one. Good for branching.
--session-id <uuid>	Use a specific UUID for the session.

Forking sessions

It's like starting claude twice in the same directory, but without using worktrees. The changes all go to the same branches and locations on the filesystem. You resume from where the most recent session left off.

claude # or if you plan ahead, "claude --name foobarbaz"
> do the foo
> /rename foobarbaz  # if you didn't set the name ahead of time

... # in another window
claude --fork-session --resume foobarbaz

I you just do --resume it'll let you pick from session history, and -c picks the newest.

When to use session forking:

• Comparison experiments. Fork the same checkpoint twice with different --model or --effort flags and see which one handles the task better.
• When you've built up context and want to try different approaches to making plans, then pick the winner to implement
• You need to /compact but you're afraid of losing something important
• Setup templating where you want to use the context like you'd use a skill/agent/md file for one-off type templated work

When to set Session ID manually

Example:

SESSION=$(uuidgen)
claude --session-id $SESSION

• Orchestrating from a script: you generate the UUID yourself, hand it to Claude, and immediately know where the transcript file will land (~/.claude/projects/<cwd>/<uuid>.jsonl). You can tail -f it, attach a log shipper, or register a webhook against it without polling the filesystem to figure out which file is "yours."
• Well-known session per job: a CI pipeline that wants "the session for PR #123" can hash the PR number into a deterministic UUID. Every job in the pipeline runs claude --session-id $uuid -c ... and they're all appending to the same transcript, reviewable in one place.
• Idempotency: if a CronJob retries, passing the same UUID resumes the prior run's conversation instead of starting a fresh one. Without it, a retry creates a brand-new session and you lose the cache + continuity.
• Dashboards / cross-system links: external UIs that want to link to a session (e.g., Linear comment → transcript) need the ID at creation time. If Claude picks the UUID, the external system has to race to discover it.
• Testing: integration tests that want predictable file paths.

Session management and viewing

Command	What it does
/rename [name]	Rename current session.
/export [filename]	Export conversation as plain text.
/copy [N]	Copy the Nth-to-last assistant response to clipboard (0 = last).
/diff	Interactive diff viewer for uncommitted and per-turn changes.
/tasks (alias /bashes)	Manage background tasks.
/recap	One-line session summary on demand.
/stats	Daily usage, session history, streaks.
/insights	Report analyzing your Claude Code sessions.
/focus	Toggle focus view (fullscreen renderer only).
/color <name>	Prompt bar color: red, blue, green, yellow, purple, orange, pink, cyan.

---

Hooks

Hooks are shell commands Claude Code runs at specific lifecycle events. You define them in settings.json. They receive JSON on stdin (including session_id, tool_name, arguments), and their exit code or JSON output can allow, block, or modify the action.

Personally, I care about hooks mostly for security and logging/notifications.

Typical uses:

• Block rm -rf or edits to .env (PreToolUse + exit 2).
• Run prettier or gofmt after every Edit (PostToolUse).
• Inject the current git branch into every prompt (UserPromptSubmit).
• Ping Discord when the session finishes (Stop).
• Gatekeep a single subagent's tool calls (hook in that agent's frontmatter).

The /hooks slash command shows what's registered and where each rule came from (user, project, managed, or plugin).

I'd originally done like a whole write-up on how to make hooks, but honestly just ask Claude to make them for you. I might give Hooks their own post.

---

CLAUDE.md and Rules

CLAUDE.md file discovery

At launch, Claude Code walks the directory tree from CWD to root and concatenates CLAUDE.md and CLAUDE.local.md from each directory. Nothing gets overridden. It all loads. During the session, when Claude edits a file it'll also load in nested CLAUDE.md files, like if Claude opens src/api/auth/login.ts, it'll pick up src/CLAUDE.md, src/api/CLAUDE.md, src/api/auth/CLAUDE.md (if any exist, but only at that moment).

Location	Scope	Shared
macOS: /Library/Application Support/ClaudeCode/CLAUDE.md	Managed org-wide	deployed by IT
Linux/WSL: /etc/claude-code/CLAUDE.md	Managed	same
Windows: C:\Program Files\ClaudeCode\CLAUDE.md	Managed	same
./CLAUDE.md or ./.claude/CLAUDE.md	Project	yes (git)
~/.claude/CLAUDE.md	Personal	no
./CLAUDE.local.md	Local	no (gitignored)

Rules - Conditionally applied CLAUDE.md

.claude/rules/ is a directory you fill with topic-scoped markdown files instead of combining everything into one big CLAUDE.md. They behave like CLAUDE.md (loaded into context at session start, treated as instructions, not enforcement) but they're organized by file.

• Files without paths load unconditionally.
• Personal rules live in ~/.claude/rules/.

To differentiate these from the usual CLAUDE.md, you need to set which paths they apply to. Add a paths frontmatter field to scope a file to specific globs:

---
paths:
  - "src/api/**/*.ts"
---
API-specific rules
foo bar baz