---
name: claude-control
description: >
  When to delegate work to a sibling Claude tab on the user's machine vs do it
  yourself, and the access differences that decide it (Hydrogen/galliaApril bypass +
  relay access vs claude.ai cloud sandbox). Trigger words: drive another claude,
  sibling claude, claude code tab, delegate to claude, run a prompt in the claude tab.
---

# Claude control (driving a sibling Claude)

Sometimes the user wants a recording or task that shows "the AI" being asked a prompt
and then doing the work. Before you wire that up, know which sibling can actually DO
the work, and that you usually can't type into it anyway.

## Which sibling can drive the user's machine?

| Sibling | Relay/bridge access? | Can drive local Fusion/browser? |
|---|---|---|
| **Hydrogen Claude** in the user's cloud env (e.g. galliaApril), bypass-permissions mode | YES (same container/relay you use) | YES, it can run the full flow autonomously |
| **claude.ai/code** web tab | NO, runs in an isolated cloud sandbox | NO, can't reach the user's local relay |

So only a Hydrogen/galliaApril-env Claude can actually open Fusion, export, drive the
browser. A claude.ai/code tab will "chug" but can't touch the user's apps.

## The catch: you usually can't type the prompt into it

The Hydrogen Claude panel input is unreachable by automation (cross-origin nested
webview, OS input no-ops), see [hydrogen-web-control](../hydrogen-web-control/SKILL.md).
So "type a 1-shot prompt into the sibling and let it run" needs the **user** to paste
the prompt (the one legitimate ask). Don't promise an autonomous sibling-driven run you
can't actually kick off.

## Decision rule

- **You have relay access too.** For almost everything, just DO the work yourself
  (drive Fusion + the browser directly), it's reliable and you control timing/captions.
  Delegating to a sibling adds a coordination + reliability tax for no gain.
- **Delegate to the Hydrogen sibling only when** the deliverable specifically needs to
  SHOW the AI being prompted (e.g. a demo of "type a prompt, AI does it"), AND the user
  pastes the prompt (or it's already running). Then you record + caption + post-produce.
- **Never run two drivers at once.** If a sibling is driving the apps via the relay,
  don't also drive them yourself, you'll collide. Pick one actor.

## If you ARE recording a sibling-driven run

CDP-driven browser actions happen in the BACKGROUND, so the recording won't show them
unless you foreground the right window as the sibling progresses. Practically, driving +
recording + captioning yourself (you control the timing) beats chasing an autonomous
sibling. See [screen-recording](../screen-recording/SKILL.md).