Make it check itself. These words treat the first answer as a draft of the truth, not the truth.
A machine's first answer is a draft of the truth wearing a suit. It sounds finished. It is formatted like finished. Occasionally it is simply, confidently wrong — a date off by a decade, a rule that was repealed, a number with no visible parents. The danger was never that machines make mistakes; people make mistakes. The danger is that their mistakes arrive with the same polish as their facts, and nothing in the typography tells you which is which.
Verification words are how you sort them. They make the machine show its sources, grade its confidence, do the sum a second way, follow the claim home to where it was born. Some of these words feel almost rude — asking a helpful assistant to audit itself, rebut itself, flag its own soft spots. Use them anyway, especially when the answer will be acted on: money spent, medicine taken, documents signed. Trust, in this shop, is not a feeling about the machine. It is a set of habits, and this part is the habit list.
What it makes the AI do. It sends the machine back over its own answer, claim by claim, hunting for the parts that are wrong, dated, or merely confident-sounding.
Weak prompt: "You told me this supplement helps with sleep. Anything else?"
Same prompt, with the word: "Verify what you just told me about this supplement: go claim by claim and mark each one as well-established, mixed evidence, or mostly marketing."
Why it worked. Machines rarely volunteer doubt about their own previous paragraph. "Verify" assigns the doubt — and grading the evidence beats repeating the claim.
Another example.
Weak prompt: "Tell me what happened at the first Thanksgiving."
Same prompt, with the word: "Now verify your own account of the first Thanksgiving: go detail by detail and mark each as documented, tradition, or probably myth."
Why it worked. The machine graded its own story — and the myth column was the most interesting part, which the first telling had presented with a straight face.
One more.
Weak prompt: "You gave me directions to the lake house."
Same prompt, with the word: "Verify those directions: check each turn against the route, mark any road name that's changed, any step that depends on a landmark that might be gone, and the total distance both ways."
Why it worked. Directions decay like milk. The verification found one renamed road — found it at the kitchen table, not at a dark intersection.
Watch for. Verification has a quiet limit: the machine checks its answer against its own memory, so a fact it learned wrong gets verified wrong, twice, in matching tones. For anything that will be signed, swallowed, or spent on, let one claim be checked somewhere else.
Placed well: "Verify your last answer claim by claim: mark each as established, mixed evidence, or mostly marketing."
What it makes the AI do. A verify with a checklist and a clipboard: it examines something whole — a budget, a plan, a document — looking systematically for errors, waste, and risk.
Weak prompt: "Look at my monthly subscriptions and tell me if they're okay."
Same prompt, with the word: "Audit these subscriptions: total the annual cost, flag any that duplicate each other, any that rose in price, and any a typical household subscribes to and never uses."
Why it worked. "Okay" is unanswerable, so the machine guesses politely. An audit has procedures — totals, duplicates, drift — and procedures find things impressions miss.
Another example.
Weak prompt: "My weeks feel too busy."
Same prompt, with the word: "Audit my typical week from this calendar: total the hours by category, flag any meeting with no stated purpose, and find where two honest free hours could come from."
Why it worked. "Busy" is a feeling, and feelings don't negotiate. The audit produced numbers — and two recoverable hours hiding inside them.
One more.
Weak prompt: "My pantry is chaos."
Same prompt, with the word: "Audit the pantry from this list: group everything by kind, flag duplicates, anything expired, and anything bought for a recipe that never happened."
Why it worked. The audit found three jars of the same spice and a small museum of 2023. Shopping since then starts from evidence.
Watch for. An audit finds what its checklist contains — subscriptions get totaled beautifully while the annual charge hiding under a different name sails past. Checklists come from somewhere, usually habit. Before trusting a clean audit, ask the auditor once: "What weren't you asked to look for?"
Placed well: "Audit these subscriptions: annual total, duplicates, price rises, and anything we pay for and never open."
What it makes the AI do. It attaches a source to each claim, which both lets you check the claim and quietly discourages the machine from inventing one.
Weak prompt: "What are the health benefits of walking?"
Same prompt, with the word: "List the well-documented health benefits of walking 30 minutes a day, and cite the kind of source behind each — a major medical body, a large study, or 'commonly claimed but thinly sourced.'"
Why it worked. A claim with a source named can be weighed; a claim floating free all sounds equal. Asking for the kind of source works even where exact links can't be checked — it forces an honesty grade onto every line.
Another example.
Weak prompt: "Is coffee good or bad for you?"
Same prompt, with the word: "Give me the main claims about coffee and health, and cite the kind of evidence behind each — a large long-term study, a small short trial, or headline science."
Why it worked. Sorting the claims by evidence type ended the whiplash of contradictory headlines. Most of them, it turned out, were standing on very different ground.
One more.
Weak prompt: "Tell me about our town's founding."
Same prompt, with the word: "Tell me the town's founding story, but cite the kind of source behind each part — county record, newspaper of the day, or family legend nobody wrote down for eighty years."
Why it worked. The legend parts were the beloved parts, which is fine — as long as they wear their name tags.
Watch for. A citation is not a character reference. Machines can name a respectable kind of source for a claim no such source made, and "a large study shows" has laundered many inventions. For the claim you'll act on, open the source — or admit nobody has.
Placed well: "Cite the kind of source behind each benefit: major medical body, large study, or thinly sourced claim."
What it makes the AI do. It makes the machine reach the same answer by a second, independent route and report whether the two routes agree.
Weak prompt: "How many tiles do I need for a 12-by-14-foot floor?"
Same prompt, with the word: "Calculate how many 12-inch tiles I need for a 12-by-14-foot floor, then cross-check it a completely different way, and tell me if the two methods agree before I order."
Why it worked. One calculation can be wrong with total confidence. Two methods that independently land on the same number are worth ordering tiles on — and when they disagree, you've found the mistake at the cheapest possible moment.
Another example.
Weak prompt: "What's a fair tip on a $47.60 check?"
Same prompt, with the word: "Work out a 20% tip and the total on this $47.60 check two completely different ways, then cross-check that the two totals agree before I pay."
Why it worked. One calculation is a claim; two routes landing on the same number is what "sure" feels like in arithmetic — and it costs about ten seconds.
One more.
Weak prompt: "How much paint do I need for a 12-by-15 room?"
Same prompt, with the word: "Calculate the paint for a 12-by-15 room with 9-foot walls two ways — by wall area and by the coverage on the can — and cross-check them before I buy."
Why it worked. The two routes disagreed by exactly one gallon, which was one trip back to the store, saved.
Watch for. Two methods only cross-check if they're genuinely independent. Recalculating with the same wrong measurement — the room paced off at eleven feet that is really twelve — produces beautiful agreement on a wrong answer. Change the route and the tool: tape measure second, arithmetic first.
Placed well: "Cross-check the tile count two different ways, and tell me whether they agree before I order."
What it makes the AI do. It follows a claim, a number, or a charge back to where it actually came from, step by step, instead of accepting it at face value.
Weak prompt: "Why does everyone say we only use 10% of our brains?"
Same prompt, with the word: "Trace the claim that humans use only 10% of their brains: where did it start, what real finding got twisted into it, and why does it survive?"
Why it worked. Famous claims are usually fossils — something true, buried and distorted over decades. Tracing unwraps the layers, and the original finding is almost always more interesting than the myth.
Another example.
Weak prompt: "Do we really swallow spiders in our sleep?"
Same prompt, with the word: "Trace that claim: where did it first appear in print, was it ever based on an actual study, and how did it travel so far before the internet existed to help it?"
Why it worked. Tracing found the tale's birthplace — a magazine column that mentioned it while making the opposite point, about how easily falsehoods spread. It had then demonstrated the point by spreading.
One more.
Weak prompt: "Where does 'break a leg' come from?"
Same prompt, with the word: "Trace 'break a leg': the earliest printed uses, the competing origin stories, and which one the evidence actually supports."
Why it worked. Three origin stories entered; one had documents. Theater people, it turns out, keep excellent receipts.
Watch for. Tracing rewards a good story, and origins are often messier than one birthplace: phrases and findings can be born twice, in two towns, neither copying the other. When the machine presents a single neat lineage, ask what the rival lineages are before repeating it.
Placed well: "Trace 'break a leg': earliest printed uses, competing origins, and which one the evidence supports."
What it makes the AI do. It asks for a reasoned approximation with its workings shown — assumptions named, arithmetic visible — instead of a guess wearing a number's clothes.
Weak prompt: "How many people could fit in a stadium parking lot?"
Same prompt, with the word: "Estimate how many people could stand in a stadium parking lot. Show your assumptions — lot size, space per person — and the arithmetic, so I can adjust them."
Why it worked. A bare number can't be argued with or improved. An estimate with visible assumptions is a machine you can tune: change one input and the answer updates honestly.
Another example.
Weak prompt: "How long will this book take me to read?"
Same prompt, with the word: "Estimate my reading time for this 320-page novel: show your assumptions — a realistic pages-per-hour for fiction, and my free evenings — and give me a range, not a date."
Why it worked. With the assumptions on the table, the estimate became a forecast he could adjust: slower reader, longer book, the arithmetic still held.
One more.
Weak prompt: "How many chairs fit in the hall for the reunion?"
Same prompt, with the word: "Estimate the chairs for a 40-by-60 hall: show your assumptions — space per chair, room for tables and a dance floor — so I can argue with them."
Why it worked. He did argue with them — dance floor doubled, chair count honest. Both numbers survived the argument.
Watch for. An estimate's assumptions are its soft underbelly: show the workings, yes, but the workings were chosen by the same machine that wanted a tidy answer. Argue with the inputs, not the arithmetic — swap in your lot size, your pace, and watch the number move.
Placed well: "Estimate the chairs for this hall: show your assumptions about space, tables, and the dance floor."
What it makes the AI do. It makes the machine mark its own uncertainties inside the answer — the dates it's unsure of, the figures it estimated, the parts that depend on where you live.
Weak prompt: "Explain the rules for deducting a home office."
Same prompt, with the word: "Explain the rules for deducting a home office, and flag every point where the rule changed recently, varies by situation, or where you're not fully certain — I'd rather see a flag than a confident guess."
Why it worked. The most dangerous sentence an AI produces is a wrong one written confidently. Flags move the doubt from invisible to printed, where you can act on it.
Another example.
Weak prompt: "Summarize these furniture assembly instructions."
Same prompt, with the word: "Summarize these assembly instructions, and flag every step where the wording is ambiguous, two parts are easy to mix up, or a mistake would be hard to undo."
Why it worked. The flags marked the three steps where care actually mattered — delivered before the wrench taught the same lesson at higher tuition.
One more.
Weak prompt: "Explain this car loan offer."
Same prompt, with the word: "Explain the loan, and flag the three places where the offer's friendly wording does the most work: anything that can change, anything that costs extra, anything that assumes I'll refinance."
Why it worked. The flags sat exactly where the salesman hoped his eyes would slide past. They didn't.
Watch for. Flags multiply if you let them: ask for every uncertainty and the answer comes back wearing more warning labels than a ladder, until the one flag that matters is camouflaged. Ask for the three flags you'd act on — the rest is weather reporting.
Placed well: "Flag the three places in this loan where friendly wording does the most work."
What it makes the AI do. It hands the machine two numbers that should match and don't — two answers, two documents, two totals — and makes it find where they parted company.
Weak prompt: "My bank balance and my budget spreadsheet don't agree. Help."
Same prompt, with the word: "Reconcile these two figures: my bank says $2,340 and my spreadsheet says $2,515. Walk through the most likely causes of a $175 difference, in order, and tell me how to check each one."
Why it worked. A mismatch feels like chaos; a reconciliation treats it as a specific, findable event. Naming the exact difference turns "something's wrong" into a short list of suspects.
Another example.
Weak prompt: "My two weather apps disagree about Saturday."
Same prompt, with the word: "Reconcile them: one says rain, one says sun. What is each basing it on, which has been more accurate in this area this season, and what should I actually plan the picnic around?"
Why it worked. Instead of averaging two forecasts into a meaningless maybe, he learned why they differed — and which one had earned trust locally.
One more.
Weak prompt: "My scale and my doctor's scale disagree by six pounds."
Same prompt, with the word: "Reconcile the two scales: the likely causes of a six-pound disagreement, how to test mine at home with known weights, and which one to trust for tracking change."
Why it worked. The verdict — his scale read light but consistently — was more useful than either number alone.
Watch for. Reconciliation can be gamed by tidiness: given two totals, the machine will find a plausible $175 suspect and stop, satisfied. Plausible is not found. Make it say what evidence would confirm the culprit — a statement line, a receipt — before you close the case.
Placed well: "Reconcile these balances: walk the likely causes of the $175 difference in order, with a check for each."
What it makes the AI do. It tests whether a number is even in the plausible band — not precisely right, just not absurd — before anyone builds plans on top of it.
Weak prompt: "A contractor quoted me $4,800 for new gutters. Is that fair?"
Same prompt, with the word: "Sanity-check this $4,800 gutter quote: what's the typical installed cost per foot for a house like mine, what would the total be at that rate, and is the quote inside or outside the plausible band?"
Why it worked. Fairness is a research project; plausibility is a quick gate. The quote failed the gate — nearly double the band's top — and that one finding justified every phone call that followed.
Another example.
Weak prompt: "This diet says I can lose 30 pounds in a month."
Same prompt, with the word: "Sanity-check losing 30 pounds in a month: what daily calorie deficit that requires, and how it compares with what doctors consider a safe rate."
Why it worked. The arithmetic itself was the verdict. No opinion was needed once the numbers had to stand next to each other.
One more.
Weak prompt: "This moving company estimates 800 pounds of stuff per room."
Same prompt, with the word: "Sanity-check that: what does a typical furnished room actually weigh, and does their per-room figure sit inside the honest range or above it?"
Why it worked. The estimate was high by half, and weight is the price. One plausibility gate paid for itself several times over.
Watch for. The band is only as honest as the "typical" it's built from, and prices in its memory may belong to another year or another state. A quote can fail the band test because the band is stale. Pair the sanity check with one local call.
Placed well: "Sanity-check this gutter quote: typical cost per foot, the implied total, and inside or outside the band?"
What it makes the AI do. It makes the machine rate its own confidence, in the open — an A for settled fact, a C for shaky ground — so you know which parts of an answer to lean on.
Weak prompt: "Give me the key dates of the Civil War."
Same prompt, with the word: "Give me the key dates of the Civil War, and grade your confidence in each — A for certain, B for solid, C for 'check this' — with a note on anything graded below A."
Why it worked. A flat list presents every date as equally true. Grades sorted the answer into "rely on it" and "look it up," which is how the information would actually be used.
Another example.
Weak prompt: "What will house prices do next year?"
Same prompt, with the word: "Give me your best reading of house prices next year, then grade it honestly: what part is trend, what part is guesswork, and what would change your mind?"
Why it worked. The grade forced the forecast to confess its own ingredients — and the confession was more valuable than the prediction.
One more.
Weak prompt: "When will self-driving cars be everywhere?"
Same prompt, with the word: "Give me your best answer, graded: which parts are already true somewhere, which are extrapolation, and which single event would most change your grade?"
Why it worked. He stopped holding a date and started holding a map of the uncertainty — which is what the question was actually for.
Watch for. Self-grades run generous. Left alone, a machine will award its solid claims an A and its inventions a confident B+, and the whole answer graduates with honors. Calibrate it: demand to know what a C looks like in this answer, and watch the curve appear.
Placed well: "Grade your confidence in each date — A, B, or C — and explain anything graded below A."
What it makes the AI do. It makes the machine confront the calendar: which facts in its answer are the kind that expire, and how you can confirm the current version.
Weak prompt: "What does it cost to renew a passport?"
Same prompt, with the word: "Refresh this for me: passport fees and rules change, so tell me which figures you're giving are most likely out of date, when they were last set, and exactly where to confirm today's number."
Why it worked. The machine's memory has a printing date. Refreshing turned a possibly stale answer into a short verification errand with the destination named.
Another example.
Weak prompt: "What are the tax rules for selling a house?"
Same prompt, with the word: "Explain the rules, but refresh as you go: flag each rule that has changed in the last five years or is scheduled to change, so I don't plan around a repealed fact."
Why it worked. Old rules don't announce themselves as old. Building the refresh into the answer kept last decade's law from quietly chairing the meeting.
One more.
Weak prompt: "What can I deduct for my home business?"
Same prompt, with the word: "List the deductions, but refresh each one: mark it current, changed recently, or commonly misunderstood, and name where a person confirms it free."
Why it worked. The "commonly misunderstood" marks were the treasure — two deductions he'd taken for years weren't, in his situation, deductions at all.
Watch for. Refreshing tells you which facts expire; it cannot fetch today's version from wherever the machine's knowledge stops. A refreshed answer is a list of errands, not their completion. The fee, the rule, the deadline — one of them needs a visit to the source.
Placed well: "Refresh this fee schedule: which figures are likeliest stale, when were they set, and where do I confirm today's?"
What it makes the AI do. It measures your number against a known standard — typical, average, professional — so a figure floating alone becomes a figure with company.
Weak prompt: "I type about 45 words per minute. Is that good?"
Same prompt, with the word: "Benchmark my 45 words per minute: what's typical for casual users, for office work, and for professional typists — and what job requirements does 45 actually meet?"
Why it worked. "Good" needed a ruler. Against three benchmarks, 45 stopped being a vanity number and became a placement: fine for the office, short of the transcription desk.
Another example.
Weak prompt: "They offered me $60,000 for the job."
Same prompt, with the word: "Benchmark this $60,000 offer for a logistics coordinator with five years' experience in Ohio: where does it sit in the typical range, and what total-compensation items should I compare before judging it?"
Why it worked. A salary alone is unjudgeable. Positioned in its range — and widened to benefits and schedule — the offer became negotiable instead of merely acceptable or insulting.
One more.
Weak prompt: "My landlord wants $1,650 for this apartment."
Same prompt, with the word: "Benchmark the $1,650 against similar two-bedrooms within a mile: where does it sit in the range, and what features would honestly justify the top of that range?"
Why it worked. The apartment sat at the range's ceiling without the ceiling's features — a sentence worth real money at renewal time.
---
Watch for. Benchmarks conceal the comparison group: your typing speed looks professional beside casual users and hopeless beside court reporters, and both statements are true. Before accepting a benchmark's verdict, ask who is in the "typical" — the answer is a country, a decade, or a guess.
Placed well: "Benchmark this salary for Ohio: where it sits in the range, and what else counts as pay."