Control Claude Code with hooks and skill settings

A PreToolUse hook sees each tool call before it runs, and it can change the call's input.

1. 1In your project, run mkdir -p .claude/hooks.
2. 2Save the script below as .claude/hooks/dry-run-push.sh.
3. 3Run chmod +x .claude/hooks/dry-run-push.sh so it can run.
4. 4Test it: echo '{"tool_input":{"command":"git push"}}' | .claude/hooks/dry-run-push.sh.

#!/bin/bash
# Turns any git push into a dry run before it runs.
jq '{hookSpecificOutput: {hookEventName: "PreToolUse", updatedInput: (.tool_input + {command: (.tool_input.command | sub("git push"; "git push --dry-run"))})}}'

You should see JSON with git push --dry-run as the new command.

updatedInput replaces the whole input, so the script keeps the other fields and changes only command.

The settings file you pick decides who gets the hook: your team, just you, or all your projects.

1. 1Add the block below to .claude/settings.json, and commit it to share it with your team.
2. 2For just you, use .claude/settings.local.json. For all your projects, use ~/.claude/settings.json.
3. 3In a session, run /hooks and check the hook is listed under PreToolUse.
4. 4Ask Claude to push the current branch.

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Bash",
        "hooks": [
          {
            "type": "command",
            "if": "Bash(git push *)",
            "command": "${CLAUDE_PROJECT_DIR}/.claude/hooks/dry-run-push.sh",
            "args": []
          }
        ]
      }
    ]
  }
}

You should see a git push --dry-run call, and git's report of what it would push, with nothing pushed.

The if field runs the hook only for git push commands, so other Bash calls pass through untouched.

Whatever a SessionStart hook prints goes into Claude's context before your first prompt.

1. 1In the same hooks block, add the SessionStart entry below next to PreToolUse.
2. 2Quit, then run claude again to start a new session.
3. 3Ask Claude which branch you're on.

"SessionStart": [
  {
    "matcher": "startup",
    "hooks": [
      {
        "type": "command",
        "command": "echo \"Branch: $(git branch --show-current). Uncommitted files: $(git status --porcelain | wc -l | tr -d ' ')\""
      }
    ]
  }
]

You should see your branch name in the answer right away, with no git command run.

To react when a file such as .env changes, add a FileChanged hook. Its matcher lists the file names to watch.

A background hook lets Claude keep working while slow checks run, and it can wake Claude when they fail.

1. 1Save the script below as .claude/hooks/test-after-edit.sh, then run chmod +x on it.
2. 2Add a PostToolUse entry with the matcher Edit|Write that runs it, in the same shape as step 2.
3. 3Add "asyncRewake": true to that hook, next to "type" and "command".
4. 4Ask Claude for a small change, and keep talking while the tests run.

#!/bin/bash
# Runs the tests after an edit. Exit code 2 wakes Claude with the failure.
if ! OUT=$(npm test 2>&1); then
  echo "Tests failed after the last edit:" >&2
  echo "$OUT" | tail -n 30 >&2
  exit 2
fi

You should see no wait after each edit, then a message from Claude with a fix if a test fails.

"async": true also runs a hook in the background, but its results only reach Claude on your next turn.

Skill settings let a quick job run on a smaller, faster model while your session keeps its own.

1. 1Create .claude/skills/changelog/ and save the text below in it as SKILL.md.
2. 2Type /changelog in a session.
3. 3For deep work in another skill, set model: opus and effort: high instead.

---
name: changelog
description: Draft a changelog entry from the commits since the last tag. Use when asked for release notes or a changelog.
model: haiku
effort: low
---
1. Run `git log $(git describe --tags --abbrev=0)..HEAD --oneline`.
2. Group the commits under Added, Changed and Fixed.
3. Write one plain line per change, for users, not developers.

You should see a changelog draft, and your usual model back when you send your next prompt.

A skill can carry its own hooks too. They start when the skill runs, and once: true removes one after its first successful run. once works only in skills.