Quick Answer
AI is fluent with numbers and unreliable at producing them, which is the worst possible combination because wrong figures arrive looking exactly like right ones. The fix is to give the model the structuring work and keep the arithmetic yourself. These 18 new prompts do that explicitly, with instructions like 'do not recalculate my numbers' and 'give me the formula, not the converted price'. The library now holds 674 prompts.
There is a category of work where AI feels most useful and is least trustworthy, and it is the work with figures in it. Forecasts, supplier quotes, weekly metrics, spreadsheet analysis, resume achievements with a percentage attached. The output arrives structured, confident and plausible, and somewhere inside it there is often a number that nobody computed.
Eighteen prompts went into the library this week to deal with that. They are not maths prompts. Almost all of them work by taking the arithmetic away from the model and giving it back to you, while letting it do the part it is genuinely excellent at: deciding what to check, how to frame it, and what is missing.
What Are the 18 New Prompts For?
They spread across nine categories, and they group into four jobs.
Checking figures before you commit to them. Sanity Check a Financial Model Before You Present It, Check a Number Before You Repeat It, Check a Dataset for Problems Before You Analyse It.
Turning raw data into a decision. Turn a Messy Spreadsheet Into an Answerable Question, Compare Supplier Quotes on More Than Price, Turn Churn Interviews into Ranked Reasons, Estimate a Project Timeline as a Range, Work Out Which Post Actually Drove the Result.
Writing about numbers honestly. Write a Chart Caption That States the Finding, Rewrite a Claim So the Evidence Is Visible, Turn a Data Table Into Prose Without Inventing Anything, Write a Weekly Numbers Note Your Manager Will Read, Write an Email That Delivers Bad Numbers Clearly, Quantify an Achievement That Has No Obvious Number.
Regional and technical specifics. Reprice a Product Across SEA Markets Without Losing Margin, Compare Marketplace Fees Across Shopee, Lazada and TikTok Shop, Write a Spreadsheet Formula and Its Failure Modes, and a classroom lesson, Teach Students to Check an AI's Arithmetic.
Every one uses labelled variables in the usual way, so you fill in your own figures rather than adapting somebody else's example.
Why Is AI Fluent With Numbers but Bad at Them?
Because writing a number and computing one are different operations, and a language model only does the first.
When a model produces a figure in the middle of a sentence, it is continuing a pattern. In context, a plausible-looking percentage is frequently the most likely next token, and plausible is not the same as correct. Nothing in the output marks the difference, which is precisely why this is dangerous in a way that a spelling mistake is not.
This is well documented rather than folklore. The GSM8K work from OpenAI, published as Training Verifiers to Solve Math Word Problems by Cobbe and colleagues, found that even the largest models of the day struggled with multi-step reasoning on grade-school arithmetic problems, despite those problems being conceptually straightforward. The team's answer was telling: rather than trying to make the model reason more reliably, they generated multiple candidate solutions and trained a separate verifier to rank them. The structural fix was to check the output, not to trust it.
The modern version of that fix is a calculator. Anthropic's code execution tool documentation describes running Python and bash in a sandboxed container specifically so Claude can analyse data and perform complex calculations rather than producing them from pattern. When a model actually runs the sum, the arithmetic is arithmetic. When it types the answer into prose, it is prediction wearing the clothes of a calculation.
Which leaves a simple practical rule. If the number was computed by code you can see, treat it as a calculation. If it appeared inside a paragraph, treat it as a claim.
How Do You Check a Figure Before You Repeat It?
Ask what would have to be true for it to be right.
Check a Number Before You Repeat It runs exactly that, and its most important line is the last one: it forbids the model from telling you whether the figure is correct. That refusal is the whole point. Ask a model whether a statistic is accurate and you will usually get a confident yes, which is worth nothing, because it has no way to know and every incentive to be agreeable.
What it can do is name the population, the time period and the definition that the figure depends on, then tell you which of those most often gets stretched when a statistic is passed from a report to a press release to a slide. In practice that identifies the problem most of the time without you ever finding the original source, because misquoted numbers usually break in one of those three places.
The same instinct sits behind Sanity Check a Financial Model Before You Present It, which opens by telling the model not to recalculate anything. Instead it attacks the assumptions: which one is carrying the forecast, which ones compound, what is missing, and which single number your board will challenge first. That is judgement work, and it is where the model earns its place.
Which Prompts Help With Spreadsheets and Data?
Three, and they run in order.
Start with Turn a Messy Spreadsheet Into an Answerable Question. It deliberately refuses to analyse anything. You give it your column names and what you are trying to decide, and it converts a vague goal into two or three questions the data can actually answer, tells you which ones your columns cannot support, and names the comparison that would make each answer meaningful. Run it before you upload the file. Deciding what you are asking costs a minute; analysing the wrong thing thoroughly costs an afternoon.
Then Check a Dataset for Problems Before You Analyse It, which builds a checking plan tuned to where your data came from. A merged export has different failure modes from a clean API pull, and the prompt asks about the specific traps: timezone shifts, cancelled rows still present, test records, currency mixing. Its final section lists the checks most likely to produce a confidently wrong answer if skipped, which is the part to read when you are in a hurry.
Then Write a Spreadsheet Formula and Its Failure Modes. It returns the formula written for your actual column letters, a plain English reading of it, what happens with blanks and text-formatted dates and inserted columns, and a test row with the answer you should expect. That last item is the one people skip. A formula returning a number is not the same as a formula returning the right number, and the test row is how you tell them apart in ten seconds.
How Do You Write About Numbers Without Inventing Any?
You forbid the model from producing figures at all.
Turn a Data Table Into Prose Without Inventing Anything is the clearest case. Every figure in the paragraph must appear in your table, no new ones may be computed, and if a comparison would help but needs arithmetic, the model must describe it in words instead of stating a calculated result. It is also told not to call anything sharp or dramatic, and not to explain why something happened, because your table contains no causes. The output is drier than what you would get otherwise and every number in it is one you can point to.
Write a Chart Caption That States the Finding fixes a smaller, extremely common problem: captions that describe the chart instead of saying what it shows. If your caption would sit unchanged under a different chart, it is a title. The prompt also flags anything in your stated finding that the chart alone would not prove, which stops the caption overclaiming.
Rewrite a Claim So the Evidence Is Visible takes a paragraph and your actual evidence, then marks each claim as supported, partly supported or unsupported before rewriting. Its most useful instruction is the one telling it not to soften everything into mush, because the standard failure mode is a model hedging claims you can fully prove alongside ones you cannot.
Two more sit at either end of the seniority range. Write a Weekly Numbers Note Your Manager Will Read forces the model to flag movements that have no explanation in your context, rather than inventing a cause for every wobble. Write an Email That Delivers Bad Numbers Clearly puts the figure in the first line and lists the two questions your reader will ask, so you prepare answers before sending rather than after.
For choosing which model to run these on in the first place, the honest comparisons at WhichAIBest cover current free tiers and context limits, and SpotFreeAI lists the tools with genuinely free plans.
What Do the SEA Pricing Prompts Do Differently?
They refuse to give you a number, and that is the feature.
Reprice a Product Across SEA Markets Without Losing Margin opens by telling the model not to produce converted prices, because exchange rates move and a model working from training data will get them wrong while sounding certain. What it gives you instead is the framework: every cost that changes when you sell into each market, the pricing conventions buyers expect in each one, how the same product gets positioned differently across the region, and the floor price formula written so you insert today's rate yourself.
Compare Marketplace Fees Across Shopee, Lazada and TikTok Shop works the same way. Fee schedules change several times a year, so the prompt builds a comparison table with the fee categories as rows and the three platforms as columns, left blank for you to fill from each seller centre. It also lists the fees sellers forget, including the ones that are optional in theory and effectively compulsory if you want traffic, plus the non-fee differences that decide profit: return rates, payout timing, cash on delivery exposure.
A blank table plus a list of exactly what to look up ages far better than any figure a model can hand you today. Our Shopee and Lazada prompt guide covers the listing side of the same platforms.
How Should You Actually Use These This Week?
Pick two. Not eighteen.
Almost everyone who works with figures repeats one checking task and one reporting task every week. Find the prompt that matches each, save both somewhere reachable in five seconds, and leave the rest until a situation calls for one. A library earns its keep by being there when a supplier sends three quotes on a Friday afternoon, not by being read end to end.
One habit is worth adopting alongside them, and it costs nothing. When a model gives you a figure, ask where it came from. If the answer is your input, use it. If the answer is the model's own knowledge, treat it as a claim needing a source. That single question catches most of what these eighteen prompts are built to prevent, and you can ask it without any prompt at all.
The library now holds 674 prompts across ten categories, all free and all usable without an account.
What Else Do People Ask?
Can AI do maths reliably yet?
Not by itself, and this is a known limitation rather than a passing bug. The GSM8K research from OpenAI found that even the largest models struggled with multi-step arithmetic on grade-school word problems, which is why the paper proposed ranking multiple candidate answers with a separate verifier instead of trusting one. Modern models handle calculation well when they are given a code execution tool and actually run the sum. Typing a figure straight into prose is the unreliable path, and it is also the default one.
Why do several of these prompts tell the AI not to calculate?
Because a model that is not allowed to produce a number cannot produce a wrong one. Instructions like every figure must appear in my table, and give me the formula rather than the converted price, move the arithmetic back to you or to a spreadsheet, while leaving the model doing structure, framing and phrasing, which it is genuinely good at. The output is less impressive and considerably more trustworthy.
What is the fastest way to check a statistic before quoting it?
Ask what would have to be true for the figure to be accurate: which population, which time period, which definition. That takes about a minute and usually exposes the problem without needing the original source, because most misquoted statistics have quietly changed one of those three things in transit. The prompt Check a Number Before You Repeat It runs this and deliberately refuses to tell you whether the figure is correct, since the model cannot know.
Should you paste a whole spreadsheet into a chatbot?
Usually not as a first move, and never with customer data unless your plan and your company policy allow it. Work out the question first, which is what Turn a Messy Spreadsheet Into an Answerable Question is for, then paste only the columns that question needs. Smaller inputs get more accurate answers, and a narrower paste is also a smaller data exposure if something goes wrong.
How many of these prompts do you actually need?
Two, most likely. Nearly everyone who works with figures repeats one checking task and one reporting task every single week. Pick the one that matches each, save them somewhere you can reach in five seconds, and ignore the other sixteen until a situation calls for one. A library is useful because it is there when you need it, not because you work through it.
Sources: Karl Cobbe et al., "Training Verifiers to Solve Math Word Problems", arXiv:2110.14168, for the finding that state-of-the-art language models struggle with multi-step mathematical reasoning on the GSM8K set of 8,500 grade-school problems, and for the verifier approach of generating multiple candidate solutions and ranking them at test time. Anthropic, Code execution tool documentation, for the sandboxed Python and bash environment used to analyse data and perform calculations rather than generating them directly. Both consulted 28 September 2026.