🏴‍☠️ A numbers whitelist caught a fake AI stat

A made-up statistic looks just as confident as a real one. That is the quiet danger of letting an AI build your slides. Nobody notices the fake number until a client or a boss asks where it came from.

The original poster on a prompt engineering forum tried something blunt. They asked an agent for a 10-slide beginner deck on pour-over coffee, then handed it a short list of the only numbers it was allowed to use: a 1:16 ratio, 15 grams of coffee to 250 grams of water, around 93 C, a 30 to 45 second bloom, about 3 minutes. Then came one line: no other numbers, with each slide capped at 40 words.

The agent came back and did something I did not see coming. It flagged that 15 grams to 250 grams is not quite 1:16, catching a rounding slip the human had written into the brief. Here is why that small moment is worth a few minutes of your attention.

Teams like Parallel, Turbopuffer, and Wordsmith are already setting the pace on Attio. Get an always-on revenue engine, with agents and workflows that build pipeline, chase every buying signal, and move deals forward with your team. Whether you're working in your browser, inbox, or favorite agent, connect to your customer data in real-time through Attio's web app, MCP, API, and SDK.

Try Attio now

An AI can hallucinate a clean-looking statistic in half a second. A fake '92% of beginners' reads with the same authority as a figure you spent an hour sourcing. That symmetry is the trap.

Think about where that bites. A made-up adoption rate in a board deck, a wrong price in a sales one-pager, a phantom benchmark in a case study: each one looks authoritative until someone traces it.

Most guardrails fail because they lean on vibes. 'Don't make things up' is a wish, not a test. What struck me here: an allowed-numbers list flips that into something you can verify line by line.

Here is why I find this worth your time: the problem is not that models lie. It is that they lie in the same font as the truth. Strip away their freedom to invent figures and you remove the single most expensive failure mode in any AI-written document.

The run used Sonnet 5.5 paired with a set of slide-building skills. One of them records where each on-screen fact came from: a step meant for decks built on attachments or research.

This deck had neither source, so the record was optional. The agent wrote it anyway, then used it as a checklist against the rule. That is the part that sold me.

It searched every slide and speaker note for numerals and found only the five approved values. It counted the words per slide, and none breached the 40-word cap. And before a single slide existed, it caught the 1:16 mismatch in the outline.

None of this was luck. The agent treated the constraint as a thing to test, not a suggestion to absorb, and the source record gave it a surface to run that test against. That distinction matters more than the coffee.

Picture the alternative. A deck full of plausible numbers nobody can trace, discovered live in a meeting. The audit turns that nightmare into a thirty-second scan at the bottom of the output.

You think 4x faster than you type. So why are you typing prompts? Wispr Flow turns your voice into ready-to-paste text inside any AI tool. Speak naturally, tangents and all, and Flow cleans it up. Available on Mac, Windows, iPhone, and Android.

Try Wispr Flow Free

_*Ad

You do not need a ten-page style guide. List every figure the agent may use, and frame approximations as starting points. Then close the door with one plain line: no other numbers.

Add a cap you can count. Words per slide, items per list, sentences per section: anything countable gives the agent a concrete target.

'Keep it short' is a wish. 'Under 40 words' is a check.

Then put the audit in the brief itself. Request a numeral search across slides and notes, plus a word count per slide, returned as a short list at the end. Don't hope it checks itself: make the evidence part of the output.

And keep the list short on purpose. A wall of rules gives the model room to interpret; a five-item list and one hard stop give it almost none. Constraint beats instruction when you want compliance you can verify.

I would push one step further and ask for the source record even when the task doesn't need it. In this run, that optional log is the exact surface the agent audited against. No log, no self-check.

Writing a constraint an agent can actually test is a knack, and it only clicks after you have watched a few briefs fail. The claude track in 3 Minute AI gives you those reps in three-minute lessons, each ending with a task you run in a built-in chat lab. Daily lessons and two full courses are free.

Connect your customer data, teams, and tools without the hassle of complex integrations or lengthy setups. One easy platform gives marketing, sales, and service teams a unified customer view and the tools to turn it into growth.

Why HubSpot and what's new
• Generate leads and automate marketing with Marketing Hub
• Build your pipeline and close more deals with Sales Hub
• Scale customer support and drive retention with Service Hub
• Keep customer data clean, connected, and actionable with one, unified platform

Join 306,000+ in over 135 countries using HubSpot to grow their businesses.

See what a more connected approach to growth can do for you and your team. Get setup quickly and start checking off your hardest tasks.

Get started free

_*Ad

Here is the catch one contributor raised: an approved number can still land in the wrong spot. A 93 meant for water temperature could surface as a bloom time, and a bare numeral search would wave it through.

The fix is simple. Pair each figure with its job. Write '93 C = water temperature' and '3 minutes = total brew time,' then check those pairs in both the slides and the notes.

This is the key: a number by itself is not a fact. It becomes one only when it is tied to what it measures, and both halves need checking.

And sanity-check your own list first. 250 divided by 15 is about 16.7, so 1:16 was always a rounded label.

Saying so up front would have saved the agent a flag.

Pull up your next agent prompt and add two lines: a short list of the exact numbers it is allowed to use, and a hard rule that says no other numbers are permitted.

If you want the exact wording, the self-audit steps, and the rounding slip it caught, read the full guardrail setup.

Worth 10 minutes if you have ever shipped a deck or a report with a number you could not actually trace back to its source.

Credits to the original creator.

Still think one invented number is no big deal? These five links say otherwise. Every story here starts the same way, with AI output that sounded right, and ends with a real person paying for it. The last one shows you how to keep it from happening to you.

Hit reply and tell us why.