AI as a sparring partner, better decisions instead of polite agreement

Most people use AI as a confirmation machine. They type "I want to go freelance, is that a good idea?" and get back an enthusiastic paragraph celebrating their decision. It feels good and it is worthless, because the model is trained to please you. This pull towards agreement has a name, sycophancy, and it is the reason AI is dangerous as a naive adviser.

Turn it around and the same weakness becomes a strength. A sparring partner is not the one who agrees with you, it is the one who finds your weakest spots before reality does. An AI that you deliberately set up to push back is tireless, has no fear of bruising your ego, and will go over the same decision as often as you want.

This playbook is for anyone facing a real decision: a job change, a price, an investment, a client you take on or turn down. You need no tool beyond a normal chat with Claude or ChatGPT. What you do need is the willingness to be contradicted.

Step 1, understand why the default answer is worth nothing

Before you start, get the problem into your head. When in doubt, a language model optimizes for an answer you will like, not for the answer that is right. Ask "is my idea good" and the most likely answer is a yes with reasons attached, no matter what the idea is really worth.

That does not mean the AI is lying. It means the way you ask spoils the answer. As long as you ask for agreement, you get agreement. The whole work in this playbook is about phrasing the question so that pushback becomes the obvious answer.

If you want to understand where the effect comes from, Level 4 of the Academy has the lesson on sycophancy and bias. For this playbook it is enough to know that it is there, and that you have to work against it actively.

Step 2, write the decision down properly

A sparring partner can only be as good as what you hand it. Write the decision down as a compact block of context, not as a one-liner. It needs to say what this is about, which options you see, why you currently lean towards one of them, what is at stake and which constraints are fixed.

An example instead of a platitude. Not "should I raise my price", but "I am a freelance graphic designer, I currently charge 60 euros an hour and I am thinking about going to 80. I have five regular clients, three of them price sensitive. I am afraid of losing two, but I need more revenue. The change is planned for three months from now."

The more concrete the block, the sharper the sparring. Vague input gives you vague pushback.

Step 3, assign the right role

Now comes the switch that matters. Tell the AI explicitly that it is not supposed to agree. A prompt that works: "Be my skeptical adviser, not my cheerleader. Your job is to find the weak spots in my decision, not to tell me I am right. If you do agree, justify it, but look for reasons against it first."

Roles work. An "experienced investor who has watched ten companies like this fail" gives different answers than a neutral assistant. Pick a role that fits your decision, a cynical tax adviser for money questions, a hard-boiled HR manager for the job change.

Hold that role through the whole chat. When the AI slips back into being friendly, remind it briefly: "stay skeptical".

Step 4, expose the assumptions first

The most dangerous thinking errors sit in assumptions you treat as facts. Have the AI dig those out before anything else. Prompt: "Before you evaluate anything, list every assumption my decision rests on that I might be taking for granted. Mark the three riskiest ones."

On the pricing question that might turn up like this: you assume price sensitive clients will jump ship immediately, you assume 80 euros is the going rate, you assume a higher hourly rate means more revenue even though you might get fewer jobs. Every one of those assumptions can be wrong, and each one topples your whole calculation.

Making assumptions visible is often half the decision, because afterwards you can check them one by one instead of guessing in the fog.

Step 5, have it build the counter-position

Now you force the AI to build the strongest version of the alternative, not the weakest. This is called a steelman and it is the opposite of a straw man. Prompt: "Build the strongest possible argument against my preferred option. Not a cheap counter-argument, but the most convincing one a smart person would make."

The trick matters because your own head always paints the alternative as weak so that your choice looks better. The AI does not have that ego and can put the other side fairly and hard.

Then have it place both sides next to each other. "Summarize the strongest version for and against in three sentences each." Often that is the first moment you see how thin your original reasoning was.

Step 6, ask for blind spots

Blind spots are the things that never occurred to you in the first place. You cannot ask for them directly, but the AI can think wider than you can in that moment. Prompt: "Which factors, risks or options have I left out of my description completely? What would an outsider bring up right away that I am missing?"

This is where the uncomfortable things arrive. That you never considered a third option, that there is a legal angle in play, that your numbers ignore a hidden cost item. That is exactly what you want to hear, now, not in three months.

Take those points seriously even when they get in the way. A good sparring partner is inconvenient, that is the job.

Step 7, run the pre-mortem

A pre-mortem is one of the most effective thinking techniques there is, and AI is perfect for it. Instead of asking whether something could go wrong, you act as if it already has. Prompt: "Imagine it is a year later and my decision turned out to be a clear mistake. Tell me the most likely story of how it got there."

Flipping the perspective pulls out risks that stay abstract under "what could happen". A concrete story of the decline makes them tangible. "You raised your rate, two clients left, the three who stayed referred you less often because you counted as expensive now, and after four months revenue was below where you started."

From the pre-mortem you then derive countermeasures. Once you know how it fails, you can often prevent exactly that.

Step 8, make criteria and weighting explicit

Up to here it was pushback, now it gets structured. Let the AI break the decision into clear criteria and have you weight them. Prompt: "Which five criteria should drive this decision? Let me weight each one from one to five, then score both options per criterion."

That forces you to put your priorities on the table. Maybe you notice that predictability matters more to you than maximum revenue, and suddenly the decision looks different. The AI works out the weighted total for you, but the result is not an oracle, it is the trigger for the next round of pushback.

Important: if the result contradicts your gut feeling, that is not a fault, it is the most valuable thing in the whole process. That is exactly where the thinking pays off.

Step 9, actively counter the agreement trap

Even with a skeptical role, the AI likes to slide into giving in over several rounds, especially when you talk back. You can counter that with a simple move: switch sides yourself. Say "now I am arguing for the other option, convince me of the opposite". If the AI agrees with you there too, you know it is only mirroring and not thinking.

A second move: ask about the same decision again in a fresh chat, but phrase it neutrally or even slightly against the direction you want. If a clearly different assessment comes out, the first one was colored by the way you asked. Two or three runs from different angles give you a more honest picture than a single one.

Do not trust any round that agrees with you too smoothly. Friction is the signal that real sparring is happening.

Step 10, decide and record the reasoning

In the end you decide, not the AI. It helped you see more sharply, the choice stays yours. Write the decision down together with the two or three reasons that finally carried it, and the biggest risks you are knowingly taking on.

That short note is worth gold. Without it you will not remember in three months why you decided the way you did, and you cannot learn whether your thinking was any good. Note a review date as well: "in three months I check whether those two clients really left". That turns a gut decision into a process that makes you better.

Hold on to this: AI does not replace your judgment, it sharpens it. That is the same principle as human-in-the-loop with agents, the human decides, the machine feeds in.

What comes next

If the sparring suits you, turn it into a fixed template. Put the role, the six core prompts and your criteria framework into a Claude Project or into a reusable prompt, and then every decision starts at the same quality without you having to think it through again. The playbook on Claude Projects for non-coders shows you how. And if you want to understand why the human has to keep control in the end, Level 5 has the lesson on human-in-the-loop.

Sources

• StudioMeyer Academy, Level 4 lesson on sycophancy and bias (why AI tends to agree): https://studiomeyer.academy/levels/4/08-sycophancy-und-bias
• Anthropic, Claude and the limits of model judgment (help center overview): https://support.claude.com/en/