skill-banner

It prints one banner the moment a skill starts, whether you typed the slash command or Claude picked the skill on its own:

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
▶ RUNNING SKILL   worktree-merge
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

It's two hooks in hooks/hooks.json and one script in scripts/banner.sh:

• When you type /name, the UserPromptSubmit hook fires, and the script reads your text as typed from the prompt field.
• When Claude picks a skill, the PreToolUse hook fires on the Skill tool, and the script reads the skill's name from tool_input.skill.

Either way the script answers with a systemMessage, and Claude Code prints it before the skill runs. Claude Code puts the hook's name in front of it (UserPromptSubmit says: or PreToolUse:Skill says:). A hook can't change that, so the box starts on the line below. A plugin skill such as claude-md-audit:claude-md-audit shows by its skill name alone, the same both ways.

Sometimes Claude fires off skills I'm totally unaware it's using. Other times, I try to fire a skill and I'm not certain if it worked or not. Claude Code starts skills two ways, and both are easy to miss. When Claude picks a skill on its own, the only trace is a tool call that scrolls past. When you type a slash command yourself, nothing confirms it took.

If you've installed more than a couple of skills, this is how you find out which ones Claude actually uses.

I had this as a CLAUDE.md rule first: "print a banner when you run a skill." A rule is text Claude reads and weighs against everything else in the prompt, including a skill that says "print nothing but the receipt." A hook runs outside the model. Claude Code runs it on the event, every time, whatever the prompt says. The hooks guide is the reference.

• Writes nothing. No files, no network calls. It needs bash, grep and sed, and nothing else. It runs on macOS and Linux.
• Never blocks Claude Code. Every exit is 0, and hooks.json wraps the script in || true.
• Asks nothing. There's nothing to customize: the banner is the whole product.
• Ignores a pasted path such as /tmp/notes.md. That isn't a command, so the script stays quiet on it.

• You type a slash command for any skill, and the banner prints before the skill's first output.
• You ask for something one of your skills covers without naming it, and the banner prints with that skill's name.
• No banner at all means the hooks haven't loaded yet. Hooks load when Claude Code starts, so quit it and start it again.

One known quirk: if a typed command's own instructions make Claude call the Skill tool for the same skill, you'll see the banner twice.