Ask for the source, not the reasoning

TL;DR
- You cannot check a chain of thought. You can check a source.
- The newer models may refuse a request to write out their reasoning.
- Ask for the source of each number, and for which checks ran and what they showed.
A few years ago, the advice was to ask the AI to "think step by step" and to show its working. It helped. The older models made fewer mistakes when they wrote their reasoning out.
That advice is now out of date, for two reasons.
The newer models do their thinking before they answer, without being asked. And a request to write that thinking into the answer can now be refused.
What changed
Anthropic's guide to its newest model says a request that "asks the model to reproduce its internal reasoning in the response text" can be refused. It goes further: "If your prompts, skills, or tool descriptions ask the model to write out its reasoning in the response, remove those instructions" (Prompting Claude Fable 5.2, as of November 2026).
Mansel Scheffel shows what such a refusal looks like in practice, and the wording that works instead, in a video from September 2026.
So if an old prompt of yours says "explain your reasoning step by step", take it out. At best it does nothing. At worst the AI declines the whole request.
Why a source is better anyway
There is a better reason to drop it. Even when the reasoning is there, I cannot do much with it.
A page of reasoning reads well. It sounds sure of itself. But I cannot check it. I can only agree with it or not.
A source I can check. If the AI says a price is $8, and says it read that on the vendor's pricing page today, I can open the page and look. If it says a number came from my own report, I can find the line.
What I want is not the AI's thinking. What I want is to know which parts of the answer are facts, where each fact came from, and which parts are guesses.
What to ask for instead
These are the requests I use now. Each one asks for something I can check.
Where did each number come from? The file and line, the page and the date, or the tool and what it printed.
Which checks did you run, and what did each one show? Not "I checked it". The check, and its result.
What is a guess? Ask the AI to say plainly which parts it could not confirm.
Anthropic's prompting page has a line I use almost word for word: "Say a check was done only if a tool result in this session showed it" (Prompting best practices, as of November 2026).
My own standing instructions now say it like this:
Every number names where it came from: the file and line, the command and what it printed, the query, or the commit. Say plainly when something is a guess.
What it looks like
Here is a report line I would not accept:
I reviewed the costs carefully and the total is about $410.
Here is one I would:
The total is $410. It is the sum of the four invoices in the folder, by the amounts on each one: $120, $95, $150 and $45. I did not check whether any were paid.
The second one is shorter on thinking and longer on facts. I can check every part of it in a minute. And the last sentence tells me exactly where to look next.
When you do want the why
Sometimes you do want to know why the AI chose something. That is fine. Ask for a short explanation of the choice, not a replay of its thinking.
"Why did you pick this layout over the other one?" is a question about the work. It gets a useful answer.
"Show me every step of your reasoning" is a request for the inside of the machine. It may be refused, and even when it is not, it is hard to use.
For a choice that is mine to make, I ask the AI to set out the trade-offs before anything is decided. I wrote about that in My favourite prompt: advise me on the trade-offs.
The other question I ask often is about the gaps I cannot see. I wrote about it in Tell me what I do not know that I do not know.
The short version
Do not ask how it thought.
Ask where each fact came from.
Ask which checks ran, and what they showed.
Ask what is still a guess.
