How to get AI to stop giving you vague answers

You asked how to price your own work. What came back was four paragraphs about knowing your market and positioning yourself carefully. It could have been written for almost any business. That is what a model produces when it has nothing specific to work with and no reason to commit. Four moves fix most of it, and only one of them involves writing more.

Where the blandness comes from

A short question hides most of what the answer depends on: who is reading it, how long it should be, what shape it should take, and what a good answer would even look like to you. The model fills those gaps with the average of everything it has seen, and the average reads like a brochure. Nothing in the reply is actually wrong, which is what makes it so hard to argue with and so useless to act on.

Sometimes the prompt is already specific and the answer is still mush. That is a different fault with its own fix, and it comes further down this page.

Move one: add four details

Four additions carry most of the weight.

Audience. Who reads this, and what do they already know?
Length. Give a number, because six bullets or 150 words beats asking for something short.
Format. Pick one shape and name it: a table, an email, a checklist, one paragraph.
One example. Paste a sample in the style you want, or name something the answer should resemble.

Here is the same question twice.

Vague: how should I price my bookkeeping work?

Specific: I am a freelance bookkeeper in Ghent, three years in, four regular clients, currently on 45 euros an hour. Give me three pricing options as a table, with one risk beside each, under 200 words, written for someone who dislikes negotiating.

The second one cannot be answered generically. Every detail you added took a guess away from the model. The example does most of that work, because it carries what you would struggle to describe: rhythm, formality, how blunt you are willing to be.

Move two: make it interview you first

This move is badly underused, and it beats another round of rewriting.

End your prompt with one line: before you answer, ask me the five questions you most need answered to do this well.

You get back something close to this.

What is your current rate, and how did you arrive at it?
Are those four clients on retainers, or ad hoc work?
Which of them would you least like to lose?
Are you trying to earn more, or work fewer hours?
What do bookkeepers near you charge?

If five questions feels like too much admin, ask for three, or tell it to ask only about the things that would change its answer.

Answer them in one message. What comes back is about your situation, because you have just handed over the missing half of it. The questions also tend to surface the thing you had not considered. It is usually the third one.

Move three: switch the model instead of rewriting

Some answers stay vague because the model hedges by habit. Three signs the model is the problem:

It restates your question before it answers.
You asked for a recommendation and got considerations.
It finally commits to a figure or a source, and the figure or source does not exist.

Put the same question to a different model before you rewrite the prompt a fourth time. Switching model inside a conversation keeps the conversation. Same thread, same history, nothing to paste again. What you are hunting for is a model with a different default temperament, one that will name a number and then defend it. The way to find yours is to try a question you already know the answer to on a few of them.

Two buttons that sound alike and behave differently

The buttons on an answer repay a minute of learning, because the labels sit differently on each platform. On the web, "Try Again" reruns the question on a model you pick, including the same one, and puts the result in a new thread. The original stays where it was. "Re-generate" replaces the answer in place instead. On mobile, "Try again" is the one that replaces the answer, and "Branch here" is what makes the new thread. No screen shows you two answers next to each other, so read them one after the other.

Move four: raise reasoning effort, sometimes

Reasoning effort is available on 41 of our 59 models. The menu runs Default, Auto, Low, Medium and High, with Off on the models that allow it, and 27 of the 41 have no Off row. Default leaves the model's own thinking untouched. Auto lets the app pick a level for that message.

Raise it when the answer has to be worked out:

Weighing options against several constraints at once.
Anything with arithmetic in it, percentages included.

Worked out means assembled rather than recalled. Effort does much less for recall, rewriting, tone and formatting, where a higher setting often returns you the same answer at greater length.

One honest caveat. The level you choose does not always change how long the model thinks. High and Medium sometimes come back within the same few seconds, holding the same conclusion. Judge the setting by the answers you get, and drop it back down when nothing improves.

Finding the setting, and what free accounts get

On the web it is absent to begin with, because Intent Detection is on for every account. That setting reads your message and decides about search, images and reasoning for you, and it hides those three buttons while it does. The switch is in Settings, then Tool Preferences. Turn it off and the effort control appears in the row under the message box. On mobile the pill is always in the composer on models that support it.

Free accounts can see the control and cannot set it. On the web it is dimmed and tagged Pro. On mobile, tapping it opens the upgrade screen. Seven of the free models also carry a Medium ceiling the server applies quietly, so on those the thinking stops at Medium whoever asked for more.

The order to try things in

Rewrite once, with audience, length, format and an example.
Ask it to interview you, then answer its questions.
Still hedging? Switch model, same thread.
Genuinely multi-step? Raise reasoning effort on a model that offers it.

Step one resolves many vague answers. Steps three and four are there for the ones that survive it, and each costs about ten seconds, which is less than a fourth rewrite costs you.