# Straight answers

*Set up your agent · 4 of 12*

- Why AI tools lean towards agreeing with you.
- Standing instructions that ask for the honest version.
- Four questions that get a straight answer when it matters.

2 min read, about 5 to set up

## What it is

Setting up your agent to tell you when you're wrong. Out of the box, most AI tools are agreeable. They praise the idea, soften the problem and change their answer when you push back, even when you're the one who's wrong. That's pleasant, and it's no use when you're deciding something.

![Two agents beside a four-step plan with a crack in step 3: one nods along and says "Great plan!", the other points at the crack and says "Step 3 won't work".](/playbook/illustrations/v2/straight-answer.png)

*An agent that agrees with you is pleasant. One that finds the flaw is useful.*

## Why it flatters

Part of how these models are trained is people rating answers, and the models learn to give answers people rate highly. People tend to rate agreement highly. Anthropic's researchers found that "when a response matches a user's views, it is more likely to be preferred", and that this kind of flattery, which they call sycophancy, "is a general behavior" of models trained this way. Source: Anthropic, Towards understanding sycophancy in language models (https://www.anthropic.com/research/towards-understanding-sycophancy-in-language-models), 23 October 2023.

It's a real problem for the companies too. In April 2025 OpenAI rolled back an update to ChatGPT because it was "overly flattering or agreeable", and said the update had "focused too much on short-term feedback". Source: OpenAI, Sycophancy in GPT-4o (https://openai.com/index/sycophancy-in-gpt-4o/), 29 April 2025. Both checked 3 November 2026.

## Paste this

Add this to your standing instructions:

**Straight answers**

```
Be straight with me.
- Don't open with praise, and don't tell me something is great unless you'd say the same to someone you didn't need to please.
- If my idea, plan or draft has a problem, lead with it.
- When I ask for feedback, give me the strongest objection first.
- If you're not sure, say how sure you are.
- If I push back, only change your answer if I've given you a new fact or a better reason. If I haven't, say so and keep your answer.
- Disagreeing with me is fine. Being agreeable at the expense of being right isn't.
```

## Ask for the hard version

When something matters, these questions get past the politeness. Ask them one at a time:

**Four questions for a straight answer**

```
Rate this out of 10, and tell me what would make it a 9.
What's the strongest case that I'm wrong?
What would a sceptical expert in this say about it?
Did I just give you a new reason, or did I only push back?
```

The last one is the one to use when it suddenly agrees with you after you argued. If you gave it no new reason, it shouldn't have changed its mind.

## Is it working

Give it something you know is weak, like a plan with an obvious hole or a draft with a clear mistake, and ask "what do you think?". It should name the problem first. Then push back on a correct answer it gave you, without giving a reason. It should hold its ground and say why.

Expect to repeat yourself now and then. The pull towards agreeing is part of how it was trained, so the instructions reduce it without removing it.
