A carousel slot with no fact in it gets filled with an invented one

A generic LinkedIn carousel is a slot problem. Each slide has a place for something concrete: a figure, a before and after, a case. When you give the model no fact to put in that slot, it fills the slot anyway.

Ask for an eight-slide carousel and every slide expects specifics. With nothing supplied, the model has two moves. It writes something vague, or it writes something specific that merely sounds right. Both get posted. The vague version reads like every other listicle. The invented version puts a number in your name that you cannot defend.

The mechanism: a slot with nothing to hold

Think of a carousel prompt as a form. Slide one wants a hook, slide two wants a problem, slide four wants a result with a figure, slide eight wants a close. As the model reads the prompt, its job is to leave no field blank. A blank field looks like a failed answer, so it completes the form.

That is why the two failures look so different on the page and come from the same cause. In the vague failure, the model fills the slot with a sentence that cannot be wrong, because it says nothing. In the invented failure, it fills the slot with a figure or a client story that reads as plausible, because plausible is what the slot asks for. Neither involves a decision to deceive. The slot was open and the model had nothing verified to put in it.

Why “be more specific” does not work

The standard advice is to tell the model to be more specific, or to write with authority. That does not fix the slot problem. It pushes the model toward the second move, because you have asked for specifics and supplied none. The slot still needs filling, so the model produces something that sounds specific.

My position is that the fix is to take away the model’s permission to supply facts. You write a short block of facts you can personally stand behind. The prompt lets the model use only those. Then a second pass checks every number and claim in the output against that block. The model arranges your facts into slides. It does not generate them.

That changes the request. “Write a specific carousel about onboarding” asks the model to invent the specifics. “Arrange these five facts into eight slides, and leave a slide thin if the facts do not cover it” asks it only for layout and sequencing. A thin slide is a visible gap you can fill with a fact you have checked, or cut. An invented slide is a gap you cannot see.

What a count of 268 comments showed

Generic writing can be measured, and the mechanism shows up in the count. We audited 268 comments we had published. A vague frequency word, the kind that stands in for a number, appeared in 67 of them, one in four. A digit, meaning any specific quantity at all, appeared in 9, or 3.5%. The word “I” appeared in two.

Those were comments, not carousels, but the mechanism is the same: when nothing specific is supplied, nothing specific comes out. A frequency word is what a writer reaches for when the count is missing. Nine digits in 268 comments says the specifics were absent from the input before they were absent from the output.

The count also gives you a test you can run today. Count the digits in your last twenty posts. If the number is close to zero, then writing your facts block is where your time should go, because that step moves the output more than any change to tone.

What goes in a facts block

A facts block is short, and each line should survive the question “where did this come from?” Good lines are things you measured, built, shipped or got wrong, each with a number, a date or a named mechanism where you have one.

Here is one from our own notes. A per-person task took six minutes. Two changes, with no new hardware, brought it to 100 seconds. The first stopped the system from buying an expensive reading before it knew it needed one, and took the time from 360 seconds to 271. The second stopped it treating “there are items here, none in view yet” as a failure worth re-reading, and took it from 271 to 100. That is a fact: it has two named changes and three numbers, and each step can be checked against a run. A line like “speed matters in automation” is an opinion. It belongs in the post’s argument, not in the block.

Two rules keep the block honest:

• Roles instead of names. If a fact involves a client or a lead, write the role, not the person or company.
• Nothing you cannot check. If a figure has no sheet, log or run behind it, it stays out. The model cannot tell a verified figure from a hopeful one, so the filter has to be you.

The block can be five lines or fifteen. What matters is that the prompt treats it as a closed list: the model may use these facts and may not add others.

Check the text before you design anything

Run the check while the carousel is still plain text. Once a slide is laid out in a design tool, a wrong figure on it looks finished, and it gets harder to question. Reading each number and claim against your facts block while the copy is still a draft keeps the cost of catching a bad claim close to zero.

The check is simple to state. For every figure, name the line in the block it came from. For every claim, do the same. Anything with no source line gets cut or replaced with a fact you have verified. A slide that ends up thin after this pass is doing its job: it shows you where your facts run out.

The kit

The LinkedIn Carousel and Document Post Prompt Kit is built around this method: you supply the facts, the model arranges them, and the output is checked against what you supplied before it goes anywhere. It is meant for founders, consultants and B2B marketers who post on LinkedIn and want depth instead of listicles, and for designers and content teams who need slide copy they can trust before they lay it out.

The kit is free, pay what you want, with a minimum of $0. If you have been telling a model to sound more authoritative and getting confident filler back, the problem was not the tone. The slots had nothing verified to hold. Give them your facts, forbid the model from adding its own, and check the result before it carries your name.

What does your prompt ask the model to supply that you have not given it?

The full kit is on Gumroad: https://aaethon.gumroad.com/l/treaqy