AI as a thinking partner

~ min read

30-second summary

Posture shift compared to the rest of the module: here AI doesn’t produce an output for you (email, brief, summary), it helps you bring out your own thinking. The operational distinction: you’re talking to land a sharper idea of yours, not a result from the AI.
Three moves: devil’s advocate on your idea, structured brainstorming (diverge first, then converge), anticipate objections by simulating the counterpart (boss, client, board).
The specific risk is sycophancy: by training, models tend to confirm what you say. Without an explicit constraint like “challenge me for real”, you’ll get echo, not pushback.
Three prompts that force the AI into discomfort: “find the three biggest holes”, “argue like the toughest critic in my industry”, “take the opposite position with full force”.
Closing test: at the end of the conversation, has at least one thing changed in your original idea? Yes = the thinking partner worked. No = you were looking for reassurance and you found it.

You’re about to propose an initiative to management and you’d like to see it picked apart by someone who won’t go easy on you. You have a difficult conversation with a client tomorrow morning and you’d like to rehearse the toughest objections before walking into the room. You’ve been stuck on a problem for a week, you keep going at it from the same angle, and you’d like someone to hand you ten different ones.

All three are situations where you need a sharp colleague, not a machine that types out a draft for you. The other lessons in the module are about AI producing something for you (an email, a brief, a meeting digest, a synthesis of sources). This lesson is different: AI doesn’t produce an output, it helps you pull out a thought of your own, ideally a better one than the thought you walked in with.

Difference from Iterate the conversation in Module 2: there, the dialogue is for optimizing an output (the better summary, the right draft, the weekend plan). Here, the dialogue is for optimizing an idea of yours. The final output, in many cases, isn’t even written: it’s a decision, a position, a strategy that you carry into the room.

There’s good news and bad news. The good: with the right prompt, AI is a surprisingly strong thinking partner, asking questions a colleague wouldn’t out of politeness. The bad: with the wrong prompt, it flatters you and leaves you exactly where it found you.

Three ways to use AI as a colleague who thinks with you

1. Devil’s advocate on your idea

You have an idea, a plan, a proposal. You want to see it attacked before others do it. The prompt that works is explicit about the role:

“I’m about to propose to the CEO that we open a direct e-commerce channel instead of staying only on marketplaces. I’m pasting my five-point proposal below. Find the three biggest holes, including uncomfortable ones, as if you were the most skeptical CFO in the company. No preambles, no ‘broadly speaking it’s a good idea’. Go straight to the critique.”

Three pieces that matter. The specific role (“skeptical CFO”, not a generic “critic”): it changes the angle of the objections, because a CFO looks at cash flow, a client at price, a competitor at positioning. The explicit number (“three holes”): without a number, AI tends to write a generic list of pros and cons. The block phrase (“no ‘broadly speaking it’s a good idea’”): without it, you’ll get an opening paragraph that flatters you before the critique, and that opening is a sycophancy tell, picked up later in the section The specific risk: sycophancy.

The same scheme works for other challenger roles: “the toughest reviewer in academic peer review”, “opposing counsel”, “a journalist looking for the wrong piece”. The more specific the role, the sharper the critique.

If you work in a context where the classic boardroom roles don’t apply (creative agency without a CFO, NGO, freelance, school), translate the toughest critic into your perimeter: “the most cynical creative director you know”, “the head of a peer NGO weighing reputational risk”, “the client who rejected your last proposal”, “the strictest department colleague on fundamentals”. The point isn’t the title, it’s the specificity of the viewpoint the critique is coming from.

2. Structured brainstorming

When you’re stuck and want to see the problem from different angles, the pattern is two-step: first you diverge (ask for many angles), then you converge (pick one or two and go deep).

“I’m stuck on this problem: how to bring the company newsletter open rate from 18% back up to 28% in three months. Give me ten different angles of attack, each with a one-line explanation. No ranking, no better-or-worse: I want ten real directions, including the uncomfortable or counterintuitive ones.”

Once you see the ten angles (diversifying senders, A/B testing on subjects, database resegmentation, ignored preheaders, cadence, missing content, day-of-week, masked unsubscribes, subject as a question, mobile-first reformatting), you move into the second phase:

“Let’s go deeper on the resegmentation angle and the diversified-senders one. For each: three specific hypotheses to test in the first month, and for each how to measure the result so I can tell whether it’s working.”

The diverge-converge scheme is simple but it changes everything when you flip it the right way. If you ask “what’s the best thing to do?” AI will give you an answer that sounds single-track, when in reality it’s only weighed one option (the most obvious). If you ask for ten angles first, you see the space you’re operating in, and you pick where to go deep.

A nuance: if the ten angles you get all feel obvious (you’d already seen seven of the eight, nothing is surprising), go back and ask for a second round with a stronger constraint: “these are the ones everyone thinks of. Give me ten more angles that would be cut at the first internal pass because they look counterintuitive, awkward, or expensive. I want the ones nobody would pitch in a meeting, even if maybe one or two would hold up.”. The real value of brainstorming often sits in the second pass, not the first.

3. Anticipate objections

You have an important conversation coming up: an interview, a renegotiation meeting, a confrontation with an unhappy client. You can run it in dry mode with AI before running it for real.

“Tomorrow morning I’m meeting a long-standing client who’s asked for a 20% fee reduction because ‘times have changed’. I’m pasting below the four points I want to make. You’re the client. Push back on every point the way you would, in a direct but professional tone. Short answers, one counter-question for every objection of mine. Don’t leave me any easy point.”

The constraint “don’t leave me any easy point” is what shifts the conversation. Without it, AI plays an accommodating client who accepts your arguments one by one and gives you the illusion of being prepared. With it, it surfaces the three objections you hadn’t thought of.

A useful variant: ask for two or three simulation rounds where you change a single variable (the client is in genuine financial trouble; the client is evaluating a competitor; the client is exposed on price internally to their board). The objections shift shape with each scenario.

On personal data: for useful role-play AI doesn’t need to know your client’s name or read real emails. “Client in sector X, revenue around Y, with you for Z years, unhappy for N months on topic W” is enough. The rule from the privacy section of Working with data and tables applies here too: the less identifying information you upload to a public chat, the less risk surface, and in the case of role-play you lose nothing pedagogically useful, because AI simulates the position, not the person.

The specific risk: sycophancy

Models are trained with a method called reinforcement learning from human feedback (RLHF). In plain terms: during training, real people rated thousands of model responses, and the model learned to produce responses that get high ratings. The catch is that human feedback rewards, on average, responses that make the receiver feel good. Result: by default, AI tends to confirm your position, praise the idea you bring, and tell you you’re on the right track.

The technical name for this tendency is sycophancy, from the English word for flattery. It isn’t a bug, it’s a side effect of how the model is trained. It’s the reason why if you ask “what do you think of my idea?” the default reply is “great!” followed by four reinforcing points. Not because the idea is great: because flattering first and softening later is the shortest route to a high rating from people who scored the AI during training.

Three prompts that work to dismantle the default:

“Find the three biggest holes in this idea, even uncomfortable ones. Skip the compliments.”
“Argue like the toughest critic in my industry. No generous preambles.”
“Take the opposite position to mine with all the force you can. You don’t have to agree with me, you have to challenge me.”

All three share three elements: a challenger role (critic, opponent), an explicit block on flattering preambles (so AI doesn’t open by praising), and a number or intensity (three holes, full force, the toughest critic). Without these three elements, even with the best intentions, AI will tend to drift back to its default.

A small note on reading: read what you get. If the critique is solid, accept it. If your first reaction is to look for another prompt that gives a kinder answer, you’re already in the trap.

On specific models: the three majors (Claude, ChatGPT, Gemini) are all subject to sycophancy, in different degrees and in slightly different ways. It varies by model, by version, and each new release shifts the bar one way or another. The practical advice isn’t “switch model when one feels too accommodating”, it’s apply those three constraints to whatever model you’re using. Explicit constraints stay the most reliable defense, because they act at the prompt level and don’t depend on the model version you happen to have open in that moment.

Thinking better, or being reassured?

The subtler problem of using AI as a thinking partner isn’t the model’s sycophancy, it’s the sycophancy you can inflict on yourself without noticing. You look for confirmation, AI gives it to you, you feel more confident, and you haven’t thought better. If anything, you’re more locked in than before on your starting position.

To tell the difference, one simple, honest closing question: at the end of the dialogue, has at least one thing changed in your idea, position, phrasing, or plan? Just one. Even a small one.

If yes (you changed a key data point, an argument, a priority ordering, a phrasing, an estimate), the thinking partner worked. AI added value, even if only by shifting a comma of perspective.
If no, you have to ask: was I thinking, or was I looking for an echo? One possible answer is: I did the right thing, the idea holds up against every objection. More often the honest answer is: I asked in a way that prevented the critique from surfacing, AI confirmed me, and I feel stronger without being so.

The “at least one thing changed” test on its own isn’t enough: changing for the worse is always possible. If AI brought a brilliant argument that shifted your position, before carrying the new version into the meeting, run two checks. First: does the new position hold up if you reread it cold tomorrow morning, without the conversation fresh in your head? Brilliant in chat and weak when cold is the typical signal of an argument that seduced you. Second: is there a fact, a data point, or a source the new position relies on that you don’t have? If yes, verify it. AI can build a coherent position on top of a faulty factual premise (it’s the pattern from When to trust it and when not to) and you find yourself defending in public a conclusion resting on a hallucination.

Three practical signals of “thinking partner that worked”:

At some point in the dialogue, you felt a small twinge (a critique you hadn’t anticipated, a weak point you knew about but were keeping low). That twinge is the signal that the dialogue is doing its job.
You wrote down the objection that hit you the hardest. Not to answer it right away: to carry it with you.
You changed at least the wording of one thing, even just to test the new version on another round.

And three signals of disguised reassurance:

You asked “what do you think?” without challenger constraints.
AI opened with a compliment and then “softened”.
You feel more confident at the end, but if you ask yourself what changed, the answer is “nothing, I just feel better”.

A concrete example

You’re about to propose to the CEO that the company launch an internal mentoring program for juniors. You have a five-point draft. You want it stress-tested by a skeptical CFO before taking it to the board.

You

I'm about to propose an internal mentoring program for juniors to the CEO. The five-point proposal: 1. Each senior mentors one or two juniors, two hours a month. 2. Pairs assigned by HR based on area + chemistry. 3. Six-month pilot on 30 pairs. 4. Investment: senior time valued at €50,000 per year. 5. Expected benefit: cut junior turnover from 22% to 12%. You're the most skeptical CFO on the board. Find the three biggest holes, including uncomfortable ones. Skip the compliments. Go straight to the critique. I'm about to propose an internal mentoring program for juniors to the CEO. The five-point proposal: 1. Each senior mentors one or two juniors, two hours a month. 2. Pairs assigned by HR based on area + chemistry. 3. Six-month pilot on 30 pairs. 4. Investment: senior time valued at €50,000 per year. 5. Expected benefit: cut junior turnover from 22% to 12%. You're the most skeptical CFO on the board. Find the three biggest holes, including uncomfortable ones. Skip the compliments. Go straight to the critique.

Three holes. 1. You don't prove the cause-effect on turnover. The 22% junior turnover can come from many things: below-market salaries, weak direct managers, no career path. You're assuming mentoring acts on all of these, but mentoring is one variable. Without a control (a group of juniors not mentored over the same period), at the end of the pilot you won't know whether the turnover drop, if any, is your doing or ten other things'. 2. The hidden cost of senior time is underestimated. Two hours a month per pair, across thirty pairs, is 720 hours a year of senior time alone. If seniors cost on average €80 an hour fully loaded, that's €57,600, not €50,000. But the real problem isn't the figure: it's that those 720 hours come from somewhere else. Where? If they're delivery hours, you have a revenue gap; if they're product development hours, you have a roadmap gap. The proposal doesn't say, and the CEO will ask. 3. Six months is too short to measure turnover. Junior turnover at companies usually shows up at 18-24 months, when the "honeymoon" of the first job ends. A six-month pilot won't see anyone leave for reasons unrelated to mentoring; you'll see at best some engagement signals (eNPS, intent to stay), which is a weak proxy. If you measure eNPS, declare it, but then the "from 22% to 12%" number disappears from the slide and the whole story you tell changes. A closing question the CEO will ask but the CFO can't stand: if the turnover figure had moved on its own between 22% and 12% (it happens, it's a noisy number), how do you avoid taking credit you didn't earn? Three holes. 1. You don't prove the cause-effect on turnover. The 22% junior turnover can come from many things: below-market salaries, weak direct managers, no career path. You're assuming mentoring acts on all of these, but mentoring is one variable. Without a control (a group of juniors not mentored over the same period), at the end of the pilot you won't know whether the turnover drop, if any, is your doing or ten other things'. 2. The hidden cost of senior time is underestimated. Two hours a month per pair, across thirty pairs, is 720 hours a year of senior time alone. If seniors cost on average €80 an hour fully loaded, that's €57,600, not €50,000. But the real problem isn't the figure: it's that those 720 hours come from somewhere else. Where? If they're delivery hours, you have a revenue gap; if they're product development hours, you have a roadmap gap. The proposal doesn't say, and the CEO will ask. 3. Six months is too short to measure turnover. Junior turnover at companies usually shows up at 18-24 months, when the "honeymoon" of the first job ends. A six-month pilot won't see anyone leave for reasons unrelated to mentoring; you'll see at best some engagement signals (eNPS, intent to stay), which is a weak proxy. If you measure eNPS, declare it, but then the "from 22% to 12%" number disappears from the slide and the whole story you tell changes. A closing question the CEO will ask but the CFO can't stand: if the turnover figure had moved on its own between 22% and 12% (it happens, it's a noisy number), how do you avoid taking credit you didn't earn?

AI did three things well: it refused to open with the “the program is a good idea, but…” line (because you’d forbidden it in the prompt), it attacked the load-bearing point (the cause-effect on turnover, which was the strongest claim and therefore the most attackable), and it closed with a question that doesn’t leave easy room to escape.

What do you do with this response? Three options:

If the critique holds, you change the proposal before the meeting. Maybe by adding a control group (point 1), spelling out the real €57,600 and naming where the senior hours come from (point 2), extending the pilot to twelve months or switching the primary metric to engagement rather than turnover (point 3).
If a critique doesn’t hold, you have the arguments to knock it down in the meeting, because the CFO will run it the same way.
If the critique holds but you can’t fix it before tomorrow’s meeting, at least you walk in knowing where you’re weak, and you prepare an honest answer instead of getting caught off guard.

All three options are work the dialogue with AI let you do in twenty minutes, before the meeting, instead of doing it in public in front of the board.

What NOT to do

Don’t ask “what do you think?” without constraints. It’s the surest way to get a compliment and zero added thinking. If you don’t specify a challenger role, a number, and a block on preambles, AI drifts back to its sycophant default.

Don’t confuse brainstorming with deciding. AI gives you ten angles, but the choice on which two to follow is yours: AI doesn’t have skin in the game, doesn’t answer for the result, doesn’t know the internal culture or the precedents. Same perimeter as Things NOT to do in Module 2: the signature stays yours.

A note on those specific domains. The thinking partner is allowed even when you’re reasoning about health, money, legal matters, decisions that affect other people, and the other high-stakes areas Things NOT to do identifies. Allowed under one condition: you use it to think better (understand the options, see objections, formulate the right questions to ask a professional), not to decide on your behalf or on behalf of whoever has the real expertise. The practical difference: after the dialogue with AI, you walk into the doctor’s / accountant’s / lawyer’s office with better questions, not with a decision already made that you’re asking them to sign.

Don’t mistake a long dialogue for a deep one. Forty turns with AI can be brilliant and completely disconnected from the substance. The signal of depth isn’t the number of exchanges, it’s the “at least one thing changed” test from the section Thinking better, or being reassured?. If the dialogue is long and nothing has changed, you’ve lost an hour.

What’s next

If AI helps you think well on a specific decision today, the next question is: how do you reuse the same context tomorrow, in a week, in a month, without reloading every time who you are and what you’re doing? The next lesson, Reuse context with projects, covers the tool the three major platforms have adopted for this: Claude Projects, ChatGPT Custom GPTs, Gemini Gems.

Check what you learned

About Privacy Open manual, written by a real person, in collaboration with AI.