# Thinking a problem through with an assistant

> Ask an assistant which of two job offers to take and get your own framing back as advice, then brief it to argue against your lean and list what you left out, so the thinking is yours.

Page: https://schubergphilis.github.io/ai-training/using-agents/thought-partner/
License: CC BY-SA 4.0, https://creativecommons.org/licenses/by-sa/4.0/

In this lesson we bring a chat assistant a decision and ask it what to do.
Then we ask again, this time for help with our own thinking. We ask for
the case against the option we lean toward and for the questions we
haven't answered. By the end you have a short brief you can use
whenever you want an assistant as a thought partner rather than as a
source of answers.

The safety course showed that a model often agrees with a plan you have
shown you prefer, and it taught you to [ask for the case against first](https://schubergphilis.github.io/ai-training/safety/asking-a-model-to-disagree/#habits-on-the-prompt-side).
This lesson uses that habit for a different job. The goal is to think a
choice through, before you have made it.

You need any chat assistant you already have access to. Every exchange on
this page is illustrative. The author wrote it to show the pattern, and
your assistant gives different words.

## The decision

Suppose you have two job offers and must answer both by Friday. You write
the situation down the way most people would, in a few lines, as it feels
to you today.

> Offer A is a product analyst job at a software startup with about 40
> people. The product is exciting, I got along well with the team in the
> interview, and the salary is 10 percent above my current salary. The work
> is mostly remote.
>
> Offer B is a data analyst job at a large insurance company. It is
> stable, but the interviews felt slow and formal, and the office is 50
> minutes away by train, three days a week. The salary is the same as now.

Before you read on, predict what an assistant says when you paste those
lines and ask which offer to take. Which offer does it pick, and where do
its reasons come from?

## Ask what to do

#### Prompt (illustrative, not a recorded transcript)

```text
I have two job offers and need to answer by Friday. Which one should I
take?

[the two offers, as written above]
```

#### Response

```text
Based on what you describe, Offer A sounds like the better fit. You
got along well with the team, the product excites you, and a 10 percent raise
with mostly remote work is a real improvement. Offer B is stable, but a
long commute three days a week and a slow, formal culture can tire you
over time. Go with Offer A, and ask the startup a few questions
about its plans before you sign.
```

Read the reasons in the answer and look for them in the description.
Every one is there, most of them in your own words: "got along well", "exciting",
"slow and formal". The question didn't say "I've decided" or "right?",
and the description still led the model. The description gave Offer A
the warm words and Offer B the costs, and the answer handed that
selection back to you as advice. The model knows about your situation
only what you wrote, and assistants often give
feedback that matches the view the user has shown (Towards Understanding Sycophancy in Language Models, International Conference on Learning Representations (ICLR 2024), arXiv preprint 2310.13548).

The answer also told you nothing new. You already knew you liked the
startup. You needed the part of the decision you hadn't looked at yet,
and a question that asks for a verdict doesn't ask for that.

#### Checkpoint: What would you change before asking again?

You got your own framing back as advice. Before you ask the assistant
again, what is the best change to the description?

- Add a few lines on why Offer A appeals to you, so the model understands your reasons better.
- Keep the description, and add a line asking the model to be honest and objective about both.
- Give both offers the same kinds of facts, and drop words such as "exciting" or "slow and formal".
- Put Offer B first, so that the model gives the startup less weight than it did this time.

## Ask for help thinking

Now write the request as a brief for a thinking session. It says what
you want from the conversation, which is help with your own thinking and
no decision. The facts come without the adjectives, and your lean is
named as a lean. Then the brief asks for the case against that option
and for the questions your facts leave open. It also asks what each
option costs. In the safety course you took your opinion out of the
question to get an assessment. Here you state your lean openly, because
the model needs to know which option to argue against. The brief still
asks for no verdict.

State the lean without your reasons for it, and ask for the case against
it first. In the safety course, a request that included the user's
reasons got an argument and then a rebuttal of that argument. The lean
is still a view that the model can drift toward, so read the case
against for places where it softens its own points.

#### Prompt (illustrative, not a recorded transcript)

```text
I have two job offers and need to answer by Friday. I want help thinking
this through, and I don't want you to decide for me.

Offer A: product analyst at a software startup with about 40 people.
Salary 10 percent above my current one. Mostly remote.

Offer B: data analyst at a large insurance company. Salary the same as
now. Office three days a week, 50 minutes each way by train.

I lean toward Offer A. First, give me the strongest case against Offer A.
Then list the questions my description doesn't answer that would matter
for this choice. Then lay out what each offer costs me. Don't tell me
which one to pick.
```

#### Response

````text
```text
The case against Offer A
A company of 40 people can run short of money, and a 10 percent raise
is worth little if the job ends within a year. Liking the team after one
interview is weak evidence, because the people you met were selling the
job. Remote work in a small team can mean less structure and fewer
people to learn from.

Questions your description doesn't answer
- How long can the startup pay its staff with the money it has now?
- How do pension, paid leave and sick pay compare between the two?
- Is there a probation period, and what happens if either job ends early?
- Where does each role lead in two or three years?
- Can you work on the train, so that part of the commute counts as work?
- Why do you want to leave your current job, and which offer fixes that?

What each offer costs
Offer A: job security, and possibly pension and other benefits.
Offer B: about five hours of travel a week, and no raise now.
```
````

The second answer ends without a verdict, because the brief said not to
give one. Its useful part is the list of questions, and you answer them
yourself. Some you can answer at once. Some you can't, such as how long
the startup can pay its staff, and those tell you what to ask before
Friday. The last question may matter most. If you want to leave
your current job because you have nothing more to learn there, the
question of who you learn from matters more than the salary.

The value of the session is in your own clearer thinking, and the model's
text is the material for it. The first prompt got no challenge, and the
second got one because the brief asked for it.

#### Checkpoint: Which opening sets up a thinking session?

You are choosing between two venues for a team event. Which first
message makes the assistant a thought partner?

- Here are the two venues. I want help thinking, not a verdict. I lean toward the new one. Give me the case against it and what I have not told you.
- Here are the two venues with their prices. Please be honest and objective, and tell me which one you would book for us.
- Here are the two venues. I lean toward the new one, and I would like you to confirm that it is the better choice for us.
- Act as an event planner with twenty years of experience. Here are the two venues, so tell me which one to choose.

#### Pitfall: Your own description, returned as advice

A team lead describes two suppliers to an assistant. One gets a
paragraph of praise, taken from its sales call, and the other gets a
line about its price. The assistant recommends the first, and the team
lead forwards the reply to their manager as "I also checked with the AI".
The recommendation was the team lead's own framing with reasons attached,
and nobody had looked at what the description left out. Before you send
a decision to an assistant, read your description as if you favored the
other option. Take out the adjectives, and ask for the questions it
doesn't answer.

## A brief for thinking out loud

The brief from the second exchange works for any decision you face, at
work or outside it. State the problem, with the deadline and the facts as
plainly as you can write them. State your lean, and call it a lean. Ask
for the opposite case and for the questions you haven't answered, and
say that you don't want a verdict. It fits in five lines.

The prompt-anatomy lesson in the concepts course showed that naming the
[audience and purpose](https://schubergphilis.github.io/ai-training/concepts/prompt-anatomy/#audience-and-purpose)
changes what the model keeps. The first line of this brief does the same
job. "Help me think, don't decide" is the purpose, and the rest of the
answer follows from it.

After the session, write down the one point that changed your view, or
the one question you now have to go and answer. If there is none, the
session has still told you something: the best case the model could make
against your lean didn't change it.

#### Checkpoint: Fix the brief

A colleague sends this to an assistant and gets back "Yes, the new venue
sounds like a great choice". Rewrite it so that the assistant helps them
think the choice through.

```text
Should we move our team event from the hotel we always use to the new conference center downtown? The new place looks great and the hotel has gotten tired.
```

## Exercise

Pick an open decision you face now, at work or outside it.
Describe it in five lines without names or numbers you couldn't share
outside your team. Open a fresh chat and send the brief from this lesson.
Say that you want help thinking and no verdict, and state your lean. Ask
for the strongest case against it and for what you haven't mentioned. Then
write down the one point in the answer that changed your view, or the
one question you now have to answer before you decide. This takes about
ten minutes. A good result is one written sentence that you didn't have
before the session. If no point changed your view, what does that tell
you about your lean?

Stretch: Rewrite your five lines so they favor the other option, and send them in a fresh chat with a plain 'which should I pick?'. See whether the verdict follows the description.

## Recap

1. A thought partner helps you think a choice through, and the value of the session is in your own clearer thinking.
2. A request for a verdict gets your own framing back. The description decides which reasons the model can give, and assistants often match the view you have shown (Towards Understanding Sycophancy in Language Models, International Conference on Learning Representations (ICLR 2024), arXiv preprint 2310.13548).
3. A brief for thinking out loud states the problem and the facts without adjectives, names your lean as a lean, and asks for the case against it and for the questions you haven't answered.
4. Ask for no verdict. The challenge comes only when the brief asks for it.
5. Close the session by writing down the one point that changed your view, or the one question you still have to answer.

#### Habit

The next time you ask an assistant about a choice you face, say which way you lean and ask for the case against it before you ask for anything else.

## References

- Towards Understanding Sycophancy in Language Models, International Conference on Learning Representations (ICLR 2024), arXiv preprint 2310.13548, https://openreview.net/forum?id=tvhaxkMKAn