Say exactly what you are saying yes to

TL;DR
- A bare "go" or a thumbs up leaves the AI to guess what you approved.
- Name the action and its details: which file, which database, which message.
- Say which routine steps need no approval at all, and say why when you say no.
An AI that works for you will stop and ask before some things. Delete this? Send that? Change the live site?
Most of us answer the way we answer a text from a friend. "Go." "Yes." A thumbs up.
To the AI, that is a puzzle. It asked about one thing, or maybe three. It has been talking for an hour. Which part did you just approve?
Name the action
Anthropic's guide to its newest model gives a plain example: "Reply 'Yes, run the migration on the staging database' where you might have sent a 'go' or a thumbs-up" (Prompting Claude Fable 5.2, as of November 2026).
The idea is simple. Say what you approve, with its details. Which action. On which thing. How many times.
"Go" approves whatever the AI thinks it was asking. "Yes, send the draft to Sarah" approves one email to one person.
Some tools now check approvals this way on their own. Claude Code has an auto mode, in which a checker reviews risky actions before they run. Its guide gives the same advice for writing rules it will read: "Name the action and its specifics" (Configure auto mode, as of November 2026).
How I do it
When my AI asks me several things at once, it numbers them. I answer by number.
Here is a real reply I sent on 20 November 2026, to a set of changes and one question:
go on 1 to 3, and 1 is yes
"Go on 1 to 3" approves three named changes. "1 is yes" answers the one question. Nothing else is approved. If the AI had a fourth idea, it is still waiting.
My standing instructions ask for questions in a set shape: what the question is, what is true today, what each answer changes, and the advised answer first. That makes my short answer safe, because the choices are already written down.
Say which steps need no approval
The other half matters just as much. If the AI asks before everything, you stop reading the questions. Then your "yes" means nothing.
Anthropic's guide puts a number on this. A vague rule about approvals made the model "hold back about half of its routine requests". Its advice: "Telling the model to act more freely is too vague, so list the routine steps that it can take on its own" (Prompting Claude Fable 5.2, as of November 2026).
So I keep two short lists.
The steps it takes without asking: run a test, read a file, save its own work as it goes, make a draft.
The steps it always asks about first: anything that sends, publishes, spends money, or changes real customer data.
Everything on the first list happens quietly. Everything on the second list comes to me as a numbered question. When a question does arrive, I know it matters.
When you say no, say why
A bare "no" stops the action. It does not stop the next one like it.
Matt Shumer makes this point in a video on the newer models (November 2026): say why when you say no.
The reason turns a single refusal into something the AI can use. "No, do not post that, it names a client who has not agreed to be named" teaches it the rule. Next time, it will check for client names before it asks.
"No" alone teaches it only that you were in a bad mood.
A checklist for your next approval
Before you send a "yes", read it back once.
Does it name the action?
Does it name the thing the action is done to?
Would a stranger reading it know exactly what I approved?
If any answer is no, add the missing words. It takes ten seconds. It can save you an afternoon of putting things back.

