The 3 prompts every PM eventually writes.

I cook the shift that runs from eleven at night to seven in the morning at Burger Xpress on Kemal Ataturk, the Banani branch, the one with the broken second flat top that we work around by running the lamb on the back burner. Between the ten o'clock dinner rush and the three o'clock drunk rush there is a long quiet hour where I sit on a stool by the prep counter and read whatever paperback survived the trip from Nilkhet that month. Lately a friend on the day shift has taken to texting me screenshots of his ChatGPT window, asking if I think the answer is right. He is a PM at some startup in Banani DOHS. He thinks I will tell him the truth because I have nothing to lose by it.

What I have noticed, watching him and the other PMs whose receipts I print at midnight, is that they all settle into the same three prompts, not five of them, and the order barely changes from one PM to the next. The trouble each one gets you into is even more fixed than that.

This is the patty everyone orders first. It comes back in the same shape every time, five tidy sections, a confident opening, and a closing line that says "by aligning stakeholders, we can drive impact" or whatever the brand's preferred filler dressed up as conviction sounds like. It looks like work the first time you read it. Then you rewrite most of it by hand. The model doesn't know which exec on the steering committee hates the word "platform," doesn't know your engineering lead just lost a sprint to a flaky deploy, doesn't know the customer in question is a regional bank that hasn't paid an invoice since November. The net effect is mildly positive in the way a faster typewriter is mildly positive, and the blank page is gone before lunch. You confuse the typewriter for a thinking partner and ship a memo that quietly argues against your own roadmap. You only notice on Thursday morning, when your director asks why you wrote that.

The second one earns its keep. Forty minutes of customer Zoom, a Slack thread that grew teeth overnight, a document that ran twelve pages someone in legal "just wanted your eyes on," and you paste it in, get the gist, move on. I use the equivalent in my own life on Tuesday evenings, when a mutton supplier sends his weekly catalogue over WhatsApp in forty consecutive messages. Summarisation has a right answer you can check against the source. You read the output, skim the source, catch the drift in thirty seconds. The risk is low, the time saved is real, and this one stays in the rotation.

This is the patty that ate my friend's Tuesday. He pasted a Jira ticket and three customer quotes and asked the model whether it was a defect to fix or a roadmap item to schedule. The model answered, confidently, in three paragraphs, and he forwarded it up to his director without reading it again. By Thursday they were arguing about a fix that didn't need to exist, and he was the one who had to walk it back to the customer.

The model will always answer, which is the whole problem. It will not say "I don't have enough context, the customer is paying you not me, go talk to them." It will match the patterns in your ticket against ten thousand others it has seen and produce a plausible verdict. Plausible is the trap. A plausible answer is exactly the kind of answer that stops you thinking, the same way a perfectly browned patty stops you from checking whether the inside is still cold and pink. You serve it. You find out at 4 AM when the customer in the orange shirt walks back to the counter.

The move I have come to use instead is to treat the model as a devil's advocate rather than an oracle. The shape of the prompt changes accordingly. You tell the model what your call is, you ask it to argue the opposite case as strongly as it can, and you have it give you three reasons you are getting this wrong. The model is now doing the thing it is genuinely good at, which is generating plausible counterarguments at speed. You are still the one weighing them at the end. The judgment stays where the org chart says it should stay. The friction in the loop is being supplied by the machine. The taste muscle you spent ten years building does not quietly slacken on a calendar nobody warned you about.

There is a Bangladeshi version of this story that I think about a lot, because my grandfather was in it. His generation of bureaucrats, the ones who ran Dhaka through the seventies from offices behind Motijheel, had a specific habit. They would hand a thick file to their peon, the office attendant who carried files and brought tea between rooms, and ask "ei ta important?" The peon would shrug or nod. That single shrug decided which file got read that afternoon and which one sat under a ceiling fan in Motijheel for six weeks. The peons were not stupid, they were just not paid to know. Over years the habit rotted how decisions got made in entire ministries, and files moved through the system on the strength of an attendant's mood that morning.

Nicholas Carr wrote a whole book about this in 2010. The Shallows is about how the tools we use to think reshape what we are able to think, and he was writing about Google and hyperlinks back when that still felt like the worst version of the problem. Every prompt is a small bet about what kind of mind you want to have in five years, and the bet is being placed faster than the mind can keep up with. The colleague of my friend who shipped the wrong fix on Thursday is now in a meeting on Monday trying to explain the bet he didn't realise he was making.

So go ahead and keep using the memo prompt and the summary prompt. They will save you a real number of hours over a quarter and your team will benefit. When the third prompt shows up, the one asking the model to make the call you are paid to make, the move is to type it if you have to and get the model's first answer out of your system. Then close the tab and answer the question yourself, the way my friend should have on the Tuesday his director ended up arguing about a fix that never needed to exist.