Most Claude Code guides stop at CLAUDE.md and slash commands. The real power is in hooks, subagents, and piping — and I haven’t seen many people write about actually using them.

Hooks — Let Claude Run Your Tests

Hooks fire shell commands in response to Claude Code events. Before a tool, after a tool, when a session ends. You configure them in ~/.claude/settings.json.

The one I use constantly — auto-run tests after any file edit:

{
  "hooks": {
    "PostToolUse": [{
      "matcher": "Edit|Write",
      "hooks": [{
        "type": "command",
        "command": "cd $CWD && bun test --bail 2>&1 | tail -10"
      }]
    }]
  }
}

Claude edits a file, hook fires, tests run, output gets injected back as context. Claude sees if it broke something and can fix it without you asking. That’s the part that felt like a mode switch for me — it stops being reactive and starts being anticipatory.

The other one I use is PreToolUse to inject reminders before Claude touches the shell:

{
  "hooks": {
    "PreToolUse": [{
      "matcher": "Bash",
      "hooks": [{
        "type": "command",
        "command": "echo 'Reminder: does this command modify production state?'"
      }]
    }]
  }
}

The hook output lands in a <system-reminder> that Claude reads before proceeding. Soft guardrail, not a hard block — but it catches things.

Subagents — Two Things at Once

Subagents let Claude Code delegate work to a separate agent instance while the main conversation keeps going. Independent work in parallel, basically.

Where I actually use this: research while planning. “Spin up an agent to read the Astro content collection docs and summarize the relevant API” — while I’m iterating on architecture in the main conversation. The subagent goes off, does the reading, reports back. No context switching from me.

If you’re using oh-my-claudecode (which I covered in the workflow post), /team and /ultrawork handle subagent spawning automatically. But you can also just ask Claude directly: “Run a background agent to explore this directory and summarize what it finds.”

The rule is independent tasks only. Subagents that share state get messy fast. But for parallel exploration or running a test suite while the main agent keeps writing? The speedup is real.

Piping — Treating Claude Like an Actual Shell Tool

The -p flag runs Claude Code headless — one prompt, stdout, exit. That’s all you need to start chaining it with other things.

# Pipe a diff, get a Slack message
git diff HEAD~1 | claude -p "Write a casual Slack update summarizing these changes for my team"

# Chain Claude with Claude
claude -p "What are the main architectural decisions in this repo?" \
  | claude -p "Turn this into a 3-bullet onboarding summary"

The git diff → Slack message one I use every few days. Two commands, no copy-paste, no manual writing. It’s a small thing, but it’s the difference between Claude Code living in a chat window versus fitting into your actual shell workflow.

The MCP integrations post covers the external connections — this is the plumbing that makes those integrations actually do something.