Ask Claude

Ask Claude

Use a full Claude session through ACP. This is not a subagent and must not be implemented with the current harness's in-chat delegation tools.

Invoke this skill only when the current user explicitly asks to run $ask-claude or explicitly says to ask Claude. Never auto-select it for review, planning, a second opinion, delegation, or implementation. Mentioning, discussing, or linking this skill is not authorization to run it.

Workflow

1. Resolve <skill-dir> to this skill's directory.

2. Give Claude a self-contained brief: objective, checkout, relevant files, constraints, expected output, whether writes are authorized, and the requirement to close its own temporary external session before finishing.

3. For advice, review, or planning, run the read-only wrapper:

   <skill-dir>/scripts/ask-claude read "<self-contained prompt>"
   

4. Use write mode only when the user authorized implementation and the brief names the owned scope:

   <skill-dir>/scripts/ask-claude write "<self-contained prompt>"
   

5. Treat the result as another full session's work. Inspect its evidence and validate any edits in the originating session before reporting completion.

6. A one-shot ACP session is temporary. Require the spawned Claude session to close itself before finishing, then verify no persistent session was left behind. The wrapper must stay on acpx ... claude exec; never switch to the persistent prompt mode unless the user explicitly requests an ongoing cross-harness conversation. On failure, cancellation, timeout, or early stop, close any exact session ID returned by ACP and report cleanup failure instead of silently leaving the session behind.

The wrapper uses the current Claude configuration; do not hard-code a model. Use a fresh one-shot ACP session by default. Continue a persistent named session only when the user explicitly asks for an ongoing cross-harness conversation. Close that named session when the user ends the conversation.

If ACP or Claude authentication fails, report the exact failure. Do not silently replace the requested Claude session with a subagent or self-answer.