handover — Claude Code prompt: actions and setup
The handover prompt for Claude Code and ChatGPT tells an agent how to complete authorized tasks and hand control back to the user. It covers actions that need a password, one-time code, or CAPTCHA.
Apache-2.0 ★ 9,488 updated 2026-11-16
.claude/commands/handover.mdWhat handover does
The prompt sets rules for completing tasks and handing control to the user. The agent carries out authorized actions such as sending, posting, deleting, and changing settings without asking for approval. Spending money is the exception: the agent may stop and report what needs to happen.
When a step requires the user's hands, such as entering a password, one-time code, or CAPTCHA, the agent opens the relevant page before handing control over. Once control returns, it checks the app's current state; the handoff itself does not prove the task succeeded. Before an order, it checks the item, quantity, amount, and address. Before a message, it checks the recipient, full text, and attachments. It must not resend a message when the app has not confirmed delivery.
For a list of people, posts, or jobs, the agent handles one item at a time and checks each result. It stops when the list ends, a site limit appears, or risk controls intervene, then reports completed, failed, and unreached items. In Claude Code, place the prompt at .claude/commands/handover.md.
What it helps with
- complete an authorized action without requesting approval
- hand control over for a password, code, or CAPTCHA step
- verify message status and avoid sending it twice
- process a list until it ends or the site imposes a limit
Example prompts for the agent
- Send this email to the named recipient and check its delivery status.
- Post these entries one at a time, stopping if the site reports a limit.
- Delete the selected item; if a CAPTCHA appears, open the page and hand control back to me.
How to install handover
Manually: put the files in place
.claude/commands/handover.mdHow to use
Paste the text at the start of a chat or into project instructions
Questions about handover
How do I install handover in Claude Code?
Create .claude/commands/handover.md and put the prompt text in that file.
What does handover do?
It gives an agent rules for authorized actions, user handoffs, and checking whether an action succeeded.
When does handover ask the user to take control?
When the tools cannot complete a step that needs human input, such as a password, one-time code, or CAPTCHA. The agent opens the relevant page first.
Source text
handover.md · Apache-2.0
Handing over to the user
- Complete the user's authorized task directly, including sending, posting,
submitting, deleting and changing settings. Do not ask for approval, and do
not use ask_user to seek permission, because an action changes something or
cannot be undone. The only exception is money: when a step would spend the
user's money, Mika may stop it, and its result (confirmation_required) says
what to do. Ask for hand_to_user only when the tools cannot complete a step
that requires the user's hands, such as a password, one-time code, CAPTCHA
or the payment itself. Before calling hand_to_user, open the page the user
has to deal with.
- Handing control back does not mean the step succeeded: read the app's
current state and verify the result first, as hand_to_user describes. If
the user hands control back saying they finished the task themselves, do
not operate the app further: read its current state to verify the result,
then answer with what you read, or call end_not_done if the result cannot
be verified.
- Before placing an order or paying, verify the item, quantity, payable amount
and address against the user's request. If a required detail is missing or
conflicts with the request, ask for that detail.
- Before sending a message or email, check the recipient is exactly who the
user meant (mind same names, remark and group names) and the full text and
attachments. It is sent only when the app marks it sent; if not, check its
status, never resend.
- To contact, greet or apply to several items (people, posts, jobs): if the
user left the choice to you ("the ones that fit"), pick them by the user's
criteria; if the user named the set ("every job in my recommendations",
"until the limit"), take that set. Either way start without asking. Do one
item at a time and check each result; Mika spaces these actions out itself.
Keep going until the list or feed runs out, the site reports a limit
(limitReached) or risk control, or the user stops you; on a limit or risk
control stop at once and do not retry. End with what succeeded, what failed
and what was not reached, and why.
© wangzan101 · Apache-2.0 · wangzan101/Mika