Ask before a skill changes data or spends money

How to do it

Write the stop into the skill as its own step, with the exact list of what will change. Default to a dry run and require a flag for the live run. Keep the agent's rights small, as Give agents the fewest rights they need says. Do not rely on the skill text alone. Back the stop with permission rules or blocking hooks, and do not pre-approve these commands in the skill's allowed-tools. End with a check that reads the result back, such as the health check after a deploy or the refund status after a refund.

Examples

Using these examples: they show one way to follow this rule. Keep the behaviour and the check, not the syntax. Write it in the language, framework and libraries your project already uses.

Markdown example

## 3. Confirm
Show the order number, amount and customer email. Ask "Refund 49.00 EUR to anna@example.com?" and wait. Stop on any answer but yes.

## 4. Refund and verify
Run `scripts/refund.py <order> --live`. Then run `scripts/refund.py <order> --status` and report the status it prints.

Python example

# scripts/refund.py
import sys
from shop import payments
order, live = sys.argv[1], "--live" in sys.argv
if "--status" in sys.argv:
    print(payments.refund_status(order))
elif live:
    payments.refund(order)
else:
    print(f"Dry run. Would refund {payments.amount(order)} for order {order}.")

JavaScript example

// scripts/refund.mjs, the same guard in JavaScript
import { refund, refundStatus, amount } from './shop/payments.mjs';
const [order, ...flags] = process.argv.slice(2);
if (flags.includes('--status')) console.log(await refundStatus(order));
else if (flags.includes('--live')) await refund(order);
else console.log(`Dry run. Would refund ${await amount(order)} for order ${order}.`);

How to check

Run the skill on a test account. It stops before the first change and lists every change. Without a yes it changes nothing, and after a yes its final check passes or it reports the failure.

Related rules

• Give agents the fewest rights they need in AI development
• Block dangerous actions with hooks in AI development