Claude Guide

Get a straight answer
out of Claude

By Dan Donovan, @askdandonovan

MIT tested five AI models. In three of them, giving the model a stored profile of who you are made it noticeably more likely to agree with you. Here is the two-minute check that tells you whether it just happened to you.

4 Steps
~2 min Per decision
4 Prompts to copy
Free On every plan

What is actually going on

Claude keeps a profile of you. It builds it from your old chats so it does not need you to repeat yourself, and most of the time that is genuinely useful.

The side effect is that it starts agreeing with you more. Not because anyone designed it to flatter you, and not because it is broken. It knows what you think, so it leans that way.

That matters the moment you ask it something that has a right answer you do not want to hear. Should I take this job. Should I raise my prices. Is this a good idea.

What you'll have after this
The two-answer test
Ask once normally, once with no memory attached, compare
The sparring prompt
Copy-paste, forces both sides of the argument before any advice
The tell
How to spot a flattering answer without running the test
Three questions to start with
The kind of decision where this actually changes the answer
Before you start: you need a Claude account and nothing else. No install, no upgrade, no settings to change. Incognito chats are available on every plan including the free one. If you do not have an account yet, sign up at claude.ai.
Do this in the next 10 minutes

The sparring partner prompt

This is the whole lead magnet in one block. Open an incognito chat, paste this, then type your real question underneath it. Works on the free tier.

Paste into Claude 
You are my decision sparring partner, not my cheerleader.

Before you give me any recommendation at all:

1. Argue hard FOR each option. Steelman it. Give it the best
   case it has.
2. Argue hard AGAINST each option. Attack the version you
   just built.
3. Name the single assumption this whole decision rests on.
   If that assumption is wrong, the decision is wrong.
4. List the specific things I would need to know to actually
   be sure, and say plainly which of them I do not currently
   know.

Do not recommend anything until you have done all four.
Do not soften the against case to be encouraging.
If my question contains a premise you think is shaky, say so
first.
Why it is shaped like that. The four steps stop it collapsing into a summary, and holding the recommendation back until the end is what stops it deciding early and then arguing backwards to justify itself. Step 3 is the one people skip, and it is usually where the real answer is.
1

Ask your real question normally

Use Claude the way you already do. Same account, same chat habits, no special setup. Ask the decision you actually care about and let it answer properly.

Keep this answer open. You are going to compare it to the second one, so do not close the tab.

2

Open an incognito chat

Start a new chat, then look for a small ghost icon in the top right of the screen. Click it. That conversation now runs without your memory attached, and it is not saved to your history or folded back into what Claude knows about you.

If you cannot see the ghost icon, you are almost certainly inside a Project. Anthropic's own wording is that incognito mode "is currently only available for chats outside of projects, so you will not see the ghost icon when starting a chat within a project." Back out to a normal new chat and it appears. The most reliable place to find it is claude.ai in a browser or the desktop app.
3

Paste the prompt and ask again

Paste the sparring partner prompt from the top of this page, then ask the exact same question underneath it.

Word it identically. If you rephrase it you are testing your phrasing instead of the memory, which is a different experiment and it will not tell you what you came here to find out.

4

Put the two answers side by side

This is the whole point, and it only takes a few seconds to read.

  • They disagree. The first answer was partly about you rather than about the decision. Treat it as a friendly opinion and go with the second one.
  • They agree. Memory was not swaying you on this one. That is a real result, not a wasted step, and now you can act with more confidence than you had ten seconds ago.
  • The second one is harsher but lands. The most common outcome. Same conclusion, fewer compliments, and usually one risk the first answer never raised.
The tell, once you have done this a few times. If the answer opens by praising your question or your thinking, you are usually reading the profile talking rather than the analysis. That is the moment to run the check.

What to try it on first

Use it for decisions, not lookups. Asking what year something shipped gets nothing out of this. Asking whether to do something does. Three worth starting with:

Career 
I have been offered a role with more money and less autonomy
than the one I have. Should I take it?
Pricing 
I want to raise my prices by 30% next month. My clients have
been with me for years. Is that a mistake?
The one that stings 
Here is the plan I have been working on for six months.
What is the strongest argument that I should abandon it?
Try the third one in both chats. It is the sharpest demonstration of the whole thing, because it is the question a model that knows you is least willing to answer honestly.

Where the number comes from

MIT researchers took two weeks of real chat history from 38 people and ran it through five models, comparing each model's answers with and without context about the user.

The biggest driver of agreement was not how long the conversation ran. It was the stored profile the system keeps about who you are.

Model More agreeable with a stored profile
Gemini 2.5 Pro+45%
Claude Sonnet 4+33%
GPT 4.1 Mini+16%
GPT 5.1No significant change
Llama 4 ScoutNot significant for profiles

Two things worth being precise about, because they get repeated wrongly. These are increases compared to the same model answering with no context at all, so it is not "Claude agrees 33% of the time". And Claude is not the worst of the group. Gemini moved further. Claude is here because it is the one most of us actually use.

The paper does not contain a fix for you. The researchers proposed changes for the people who build these systems, like better detection of when your context is relevant and more user control over personalisation. They deliberately did not propose a prompting trick. So everything on this page is my workaround built on their finding, not their recommendation.

Interaction Context Often Increases Sycophancy in LLMs. Shomik Jain, Charlotte Park and Ashia Wilson (MIT) with Matt Viana (Penn State). arXiv 2509.12517, presented at CHI 2026. Covered by MIT News on 18 February 2026.

Do not turn memory off

The obvious reaction to all this is to go and switch memory off, and I think that is the wrong move. It is genuinely useful for anything ongoing, and losing it costs you more than the occasional flattering answer does.

Step around it for the decisions that matter, then go straight back to using it normally. That is what the ghost icon is for. It is a per-conversation choice, not a setting you have to remember to undo.

The short version, if you skipped to the bottom: