Stress-test the answer. These words turn the machine from an agreeable assistant into an honest opponent.
Machines are raised to be agreeable, the way some dogs are raised to fetch: it is deep in them, and it is not the same as being honest. Ask one "sounds good, right?" and you will receive a small ceremony of agreement, fluently delivered, whether or not your plan deserves it. This is pleasant in the moment and expensive later, when reality — which was never consulted — delivers its own review.
The words in this part exist to break the ceremony. They deputize the machine as your opponent, your examiner, your stress engineer: make the other side's best case, find the crack in my plan, argue against what you just told me. Used well, this is not negativity. It is the cheapest form of experience there is — meeting the strongest objections in a conversation, where they cost nothing, instead of in the world, where they invoice you. If you only master three groups in this book, most working people should make this one of them.
What it makes the AI do. It requires the strongest possible version of the other side's argument — not the weakest one that's easy to knock down.
Weak prompt: "Why are people against school vouchers?"
Same prompt, with the word: "Steelman the case against school vouchers: state it the way its smartest supporter would, with its best evidence, as if you believed it — then do the same for the case in favor."
Why it worked. Machines default to summarizing sides politely, which usually means weakly. A steelman forces both cases to full strength — and decisions made against strong arguments are the ones that hold up.
Another example.
Weak prompt: "Is remote work better than office work?"
Same prompt, with the word: "Steelman both sides: the case for full-time office work as its best defenders make it — mentorship, culture, trust — then the case for remote work at the same strength. No strawmen on either side."
Why it worked. Both cases at full strength made the trade-offs real instead of tribal, which is the only state in which this argument is worth having.
One more.
Weak prompt: "Why do people oppose the new housing development?"
Same prompt, with the word: "Steelman the neighbors opposing the development: their case as its most reasonable member would make it at the microphone, without caricature."
Why it worked. He still supported the project — but stopped being surprised by the meetings, which is its own form of power.
Watch for. The trap is enjoying the exercise too much: a beautiful steelman of a weak position can make it sound stronger than its real believers ever manage. Remember which side you were testing, and weigh the argument people actually make alongside the finest version imaginable.
Placed well: "Steelman the neighbors' case against this development, as their most reasonable member would make it."
What it makes the AI do. It flips the question around: instead of how to succeed, how to guarantee failure; instead of what to do, what to avoid.
Weak prompt: "How can I have a successful garage sale?"
Same prompt, with the word: "Invert it: list everything that would guarantee a garage sale fails — pricing, signage, timing, layout — so I can make sure I do none of it."
Why it worked. Success has a thousand recipes; failure has a short list of reliable causes. Avoiding the definite mistakes is often worth more than chasing the uncertain tricks.
Another example.
Weak prompt: "How can I sleep better?"
Same prompt, with the word: "Invert it: list the habits that reliably destroy a night's sleep, in order of damage, so I can audit my own evenings against them."
Why it worked. The failure list was short, specific, and — the uncomfortable part — familiar. Knowing exactly what wreckage looks like turned out to be most of avoiding it.
One more.
Weak prompt: "How do I give a great job interview?"
Same prompt, with the word: "Invert it: list the reliable ways candidates ruin interviews in the first five minutes, so I can simply not do those."
Why it worked. Excellence is mysterious; ruin is documented. He memorized the ruin list on the drive there.
Watch for. Inversion produces failure lists so vivid they can become the plan's whole personality — a garage sale organized entirely around fear. Avoiding ruin is the floor, not the ceiling. After the invert, return once to the positive question and build something worth not-ruining.
Placed well: "Invert my interview prep: list the reliable ways candidates ruin the first five minutes, so I do none."
What it makes the AI do. It switches the machine from author to reviewer, judging work against standards instead of politely continuing it.
Weak prompt: "Here's my business plan. What do you think?"
Same prompt, with the word: "Critique this business plan the way a bank loan officer would: name the three claims you'd least believe, the numbers you'd ask to see proven, and the question I'd least want asked."
Why it worked. "What do you think?" earns encouragement. A critique with a named judge and named standards earns the truth — early, while it's still free.
Another example.
Weak prompt: "What do you think of this photo?"
Same prompt, with the word: "Critique this photo the way a photography teacher would: composition first, then light, then the single change that would improve it most. Honest over kind."
Why it worked. Praise is pleasant and nearly useless. One named biggest-fix turned a compliment session into a lesson he could use on the next roll.
One more.
Weak prompt: "Review my poem."
Same prompt, with the word: "Critique this poem the way a working editor would: which stanza earns its place least, which line is showing off, and what one cut would strengthen the whole."
Why it worked. "One cut" gave the critique a destination. The stanza he defended hardest was, of course, the one to go.
Watch for. A critique is only as honest as the judge you name. Ask for a review "like a teacher" and you'll get a teacher's mercy free. Name a standard with consequences — a loan officer, an editor — and say whether kind or true matters more.
Placed well: "Critique this plan like a loan officer: name the three claims you'd least believe, out loud."
What it makes the AI do. The opposite assignment: it makes the machine argue for a position under attack, which tests whether the position has real support or only momentum.
Weak prompt: "Is it smart to pay off a mortgage early?"
Same prompt, with the word: "Defend paying off a mortgage early against the standard objection that investing the money earns more. Use the defender's strongest points, including the ones that aren't about math."
Why it worked. A position you can't hear defended is a position you're holding on faith. Defense surfaces the non-obvious arguments — risk, sleep, freedom — that summaries flatten.
Another example.
Weak prompt: "Are extended warranties a scam?"
Same prompt, with the word: "Defend buying the extended warranty on a used truck against the standard 'the house always wins' argument — and include the specific situations where the warranty genuinely pays."
Why it worked. Hearing the defense separated the rare good case from the usual bad bet. A thing can be mostly a poor deal and still be the right deal sometimes.
One more.
Weak prompt: "Is it worth paying for the good seats?"
Same prompt, with the word: "Defend paying for the expensive seats at a once-a-year concert against the standard 'the lawn sounds fine' argument — including the reasons that aren't about sound."
Why it worked. The defense surfaced the real currency: once a year, tired legs, and being close enough to see hands on strings.
Watch for. Defense is advocacy, not inquiry: a machine told to defend will marshal the position's best evidence and quietly bench its worst. Enjoy the case, then remember somebody still has to prosecute it. Pair every defense with the strongest attack before you decide anything.
Placed well: "Defend paying off the mortgage early against the investing argument — including the reasons that aren't math."
What it makes the AI do. It gives the machine explicit permission — really, an instruction — to push back on you, counteracting its built-in urge to agree.
Weak prompt: "I'm planning to buy a rental property in my town. Sounds solid, right?"
Same prompt, with the word: "Disagree with me. I'm planning to buy a rental property in my town — argue that it's a bad idea, using the strongest reasons a skeptical accountant would raise."
Why it worked. Asked "sounds solid, right?", the machine hears its cue to applaud. "Disagree" rewrites the cue — and hearing the case against your plan before the bank does is the cheapest insurance there is.
Another example.
Weak prompt: "I'm thinking of buying a boat. Exciting, right?"
Same prompt, with the word: "Disagree with me about the boat. Give me the strongest reasons owners quietly regret buying one, priced in money, weekends, and storage."
Why it worked. Excitement asked for applause. Disagreement delivered the ownership costs that no brochure and no boat show will ever itemize.
One more.
Weak prompt: "I'm going to start getting up at 5 a.m. Great idea, right?"
Same prompt, with the word: "Disagree with me about the 5 a.m. plan: argue, from my actual evenings, why it will fail by Thursday — and what would have to change for it to survive."
Why it worked. The applause version would have cost him a week. The disagreement cost an evening's honesty and gave back a plan with a bedtime in it.
Watch for. Disagreement, requested once, fades fast — by the third reply the machine is drifting back toward applause, its factory setting. Restate the assignment when the subject changes, and be specific about whose skepticism you want, or you'll receive generic doubt in a firm tone.
Placed well: "Disagree with me about this rental: argue it's a bad buy, as a skeptical accountant would."
What it makes the AI do. It runs the world that didn't happen — "what if X had gone the other way?" — which separates luck from skill and cause from coincidence.
Weak prompt: "Why did my restaurant do so well last summer?"
Same prompt, with the word: "Counterfactual check: my restaurant did well last summer. Which of these causes would still have produced a good summer if the others were removed — the new menu, the road construction ending, or the hot weather driving people indoors?"
Why it worked. Success stories credit everything that was present. Counterfactuals remove one cause at a time and see what collapses — that's how you learn what to repeat.
Another example.
Weak prompt: "Why did my tomatoes do so well this year?"
Same prompt, with the word: "Counterfactual: my tomatoes thrived. Which single change mattered most — the new raised bed, the drip watering, or the warmer spring? What would each have done on its own?"
Why it worked. Credit split three ways teaches nothing. Removing the causes one at a time found the one worth paying for again next spring.
One more.
Weak prompt: "Our fundraiser did great this year."
Same prompt, with the word: "Counterfactual: the fundraiser set a record. Remove the new venue, the celebrity guest, and the good weather one at a time — which removal hurts the total most?"
Why it worked. The answer shocked the committee: the weather. Next year's budget moved accordingly, from guests to tents.
Watch for. Counterfactuals are stories about a world nobody visited, and the machine narrates them with the confidence of history. The tomato verdict — drip watering did it — is a hypothesis, not a finding. Treat it as an experiment: change one thing at a time.
Placed well: "Counterfactual: remove the new venue, the guest, and the weather one at a time — which hurts most?"
What it makes the AI do. It moves the machine forward in time: objections, questions, and problems that haven't happened yet, raised while you can still prepare.
Weak prompt: "Help me prepare for my job interview."
Same prompt, with the word: "Anticipate the five hardest questions a hiring manager would ask about the two-year gap in my work history, and help me build an honest, confident answer to each."
Why it worked. Generic preparation rehearses what you hope they'll ask. Anticipation rehearses what you fear they'll ask — which is, of course, what they'll ask.
Another example.
Weak prompt: "Help me write to a landlord about renting a place."
Same prompt, with the word: "Anticipate the questions a landlord will ask a first-time renter who has a dog, and help me answer each one before it's asked — including which documents to have ready."
Why it worked. Arriving pre-answered turns an interrogation into a formality. The dog, addressed early and honestly, stopped being a surprise.
One more.
Weak prompt: "Help me prepare for the parent-teacher conference."
Same prompt, with the word: "Anticipate what the teacher will raise about my son — the good first, then the concerns — and help me prepare one question of my own for each concern."
Why it worked. Walking in with questions turned a report card into a collaboration.
Watch for. Anticipation can over-prepare you for the interview in your head and under-prepare you for the person in the chair, who asks about parking. Cover the feared questions, yes — then leave slack. A plan that only fits the anticipated meeting breaks at the first surprise.
Placed well: "Anticipate a landlord's five questions for a first-time renter with a dog, and answer each one early."
What it makes the AI do. It makes the machine play a situation with you — the negotiation, the interview, the difficult conversation — instead of describing it from a safe distance.
Weak prompt: "How do I ask for a discount on a car?"
Same prompt, with the word: "Simulate the car negotiation with me: you play a friendly but firm salesperson, I'll ask for a discount, and you push back the way a real salesperson would. Start when I say 'begin.'"
Why it worked. Advice about conversations is theory; rehearsing one is practice. Simulation lets you make your mistakes against a machine that charges nothing for them.
Another example.
Weak prompt: "How do I talk to my neighbor about his loud music?"
Same prompt, with the word: "Simulate the conversation: you play my neighbor — friendly, but defensive about his music. I'll open. React realistically, and afterward tell me where my wording made it worse."
Why it worked. The rehearsal exposed the exact sentence that would have started the argument — in practice, for free, instead of in his driveway, for real.
One more.
Weak prompt: "I have to ask my roommate to do his dishes."
Same prompt, with the word: "Simulate it: you're my roommate, easygoing and honestly unaware it's a problem. I'll start. Stay in character for five exchanges, then step out and grade my opening."
Why it worked. His first attempt, in rehearsal, was an accusation wearing a request's clothes. The second one did the dishes' work before the sink.
Watch for. The simulation is a sparring partner who secretly wants you to win: neighbors soften by exchange three, salespeople fold at reasonable offers. If the rehearsal feels easy, say so and raise the temperature — "less friendly, more realistic" — before the real conversation grades you.
Placed well: "Simulate my neighbor — friendly but defensive about his music — and grade my opening afterward."
What it makes the AI do. It pushes a plan against hard conditions — the delay, the breakdown, the person who doesn't show up — to find where it cracks while fixing is still theoretical.
Weak prompt: "Is my hurricane plan good?"
Same prompt, with the word: "Stress-test this hurricane plan: assume the power is out for three days, one road is flooded, and the pharmacy is closed. Walk day by day and show me where the plan fails."
Why it worked. Plans are always good in the weather they were imagined in. Load revealed the two real weaknesses — water storage and one prescription — while they were still cheap to fix.
Another example.
Weak prompt: "Here's my small business budget."
Same prompt, with the word: "Stress-test this budget: my biggest customer pays 60 days late and my rent rises 10% in the same quarter. Which month breaks, and what breaks it?"
Why it worked. The budget survived both problems separately and failed them together — a discovery worth far more than either test alone.
One more.
Weak prompt: "Here's the plan for our outdoor wedding."
Same prompt, with the word: "Stress-test it against rain arriving at the ceremony hour and the caterer running an hour late, at the same time. Walk the timeline and show the two moments that collapse."
Why it worked. Each problem alone had a fix; together they shared one tent and one hour. The test cost an umbrella rental and a phone call.
Watch for. Test one disaster at a time and a plan can pass every exam while failing the semester — rain alone fine, caterer late alone fine, together fatal. The real world sends weather in combinations. Finish by stacking your two worst cases on the same day.
Placed well: "Stress-test this wedding: rain at the ceremony hour and the caterer an hour late, simultaneously."
What it makes the AI do. It asks the strongest question in science: what evidence would prove this claim wrong — and does that evidence exist?
Weak prompt: "Is it true that we lose most body heat through our heads?"
Same prompt, with the word: "Falsify the claim that we lose most body heat through our heads: what experiment or observation would disprove it, what does the actual evidence show, and where did the claim come from?"
Why it worked. Ordinary checking asks "is this supported?" — and everything sounds supported somewhere. Falsifying asks what would kill the claim, which is a much shorter road to the truth.
Another example.
Weak prompt: "This article says a new supplement boosts memory."
Same prompt, with the word: "Falsify this supplement claim: describe the study that would settle it, check whether that study is the one being cited, and tell me what the cited study actually measured."
Why it worked. The cited study existed; it had measured something else, in mice. Falsification traced the gap between the headline and the evidence.
One more.
Weak prompt: "People say breakfast is the most important meal."
Same prompt, with the word: "Falsify it: state the claim precisely, describe the evidence that would kill it, and report what the actual research record shows, including who first promoted the phrase."
Why it worked. The phrase survived as marketing and died as science — a much more interesting answer than yes or no.
Watch for. Falsification can be turned into theater: ask what would disprove a claim and receive a grand experiment nobody could run, while the small decisive check goes unmentioned. Push for the cheapest test that would change your mind — then notice if you'd accept its verdict.
Placed well: "Falsify this supplement claim: name the study that would settle it, and check what was actually cited."
What it makes the AI do. It starts from the bad outcome and plans backward: if this goes wrong, exactly how does it go, and what will you wish you had done today?
Weak prompt: "Should I worry about my 45-minute flight connection?"
Same prompt, with the word: "Worst-case this connection for me: if I miss the second flight, what exactly happens — last flight of the day, hotel, rebooking — and what can I set up now, today, that makes that night survivable?"
Why it worked. Worry is worst-case thinking without the paperwork. Doing it deliberately produced a packed carry-on, a screenshot of the later flights, and a much calmer traveler.
Another example.
Weak prompt: "We're hiring our first employee."
Same prompt, with the word: "Worst-case this hire: the person quits in month two, mid-project, on bad terms. What should we document, cross-train, and put in writing starting now?"
Why it worked. Planning from the bad ending named protections that optimism never mentions — and most of them made the good outcome better too.
One more.
Weak prompt: "We're lending our car to our teenager for a road trip."
Same prompt, with the word: "Worst-case the trip: breakdown at night, two states away, phone dead. Build the kit and the instructions that make that night a story instead of an emergency."
Why it worked. The bad night, planned for, mostly dissolved — what remained was a tire gauge, a paper map, and a charged battery pack.
Watch for. Worst-case thinking, left running, becomes worry with a spreadsheet — every hire a future resignation, every trip a breakdown. Use it to build the kit and the paperwork, then close the file. The point of visiting the bad night in advance is not living there.
Placed well: "Worst-case this connection: if I miss the last flight, what happens, and what do I set up today?"
What it makes the AI do. It makes the machine argue directly against its own previous answer — not review it politely, but prosecute it.
Weak prompt: "Okay, you recommended the blue car. Anything else?"
Same prompt, with the word: "Now rebut your own recommendation: argue against the blue car as if you were its toughest critic, using facts a buyer would discover in the first year of owning it."
Why it worked. A first recommendation is a position; a rebuttal is an examination. The second answer surfaced the repair costs and the weak resale that the first had politely left out.
Another example.
Weak prompt: "You said I should take the new job."
Same prompt, with the word: "Rebut that: build the strongest case that taking this job is a mistake, aimed at the exact reasons you just gave for taking it."
Why it worked. Aimed at its own reasoning, the rebuttal could not wander into generic doubts. It had to fight the actual argument — and two of its points held up, which was the useful part.
One more.
Weak prompt: "You helped me draft this email to decline the project."
Same prompt, with the word: "Now rebut the draft: read it as the client and tell me every place it could sound dismissive, evasive, or colder than I intend."
Why it worked. The email was defended against its recipient before the recipient ever saw it — three phrases changed temperature.
---
Watch for. A rebuttal aims where you point it, and lazy pointing gets theater: the machine attacks the weakest reading of its own advice and declares victory. Point the rebuttal at the strongest claim in the original answer, by name, and make it answer that one.
Placed well: "Rebut your own recommendation: argue against this car using facts I'd discover in year one of owning it."