Some AI Skills Make Your Model Dumber. Here's How to Spot Them.

Most advice about AI skills treats them as uniformly harmless. Worst case, the thinking goes, a bad skill just doesn't do much. Add it and maybe nothing improves, but nothing breaks either. Free upside.

That's wrong, and the way it's wrong is expensive.

Some skills can only ever give you a mediocre answer. Others can hand you a confidently wrong one — and you won't catch it, because it arrives in the same authoritative voice as everything else the model says. The second kind actively degrades the model's judgment on the exact topic you consulted it about. It makes the model dumber than it was before you "helped" it.

The good news is that the two kinds are easy to tell apart once you know what separates them. Here's the split, and here's how to spot the dangerous one before it costs you.

Two kinds of skills

Nearly every skill is one of two types.

A procedure or voice skill shapes how the model works: the steps it follows, the format it produces, the tone it writes in. "Sort this brain dump into three priorities." "Format the output as a scorecard." "Write like a plainspoken mentor, no corporate filler." These skills don't tell the model new facts about the world. They tell it how to behave.

A knowledge or context skill injects what the model treats as true: facts, figures, rules, the specifics of your process or your industry. "Our fiscal year starts in February." "Here's how the 2026 LinkedIn algorithm ranks posts." "This is the exact reconciliation procedure for our books." These skills change the model's picture of reality before it answers.

That distinction looks academic. It isn't. It's the difference between a skill that fails safe and one that fails dangerous.

Procedure skills fail safe

When a procedure skill is bad, the damage is capped. The model still reasons the way it always does; it's just following clumsy instructions about format or sequence. You get an awkwardly structured answer, or a tone that misses, or a step in the wrong order. You notice, because the output looks off. You adjust the skill or ignore it.

Nothing about a weak procedure skill corrupts the model's actual thinking. It's a bad set of stage directions handed to a capable actor. The performance suffers; the actor doesn't forget how to act. Worst case: mediocre. You can see mediocre coming, and you can route around it.

Knowledge skills fail dangerous

Knowledge skills are a different animal, because of one thing the model does that most people never think about: it keeps a running sense of how sure it is.

Ask a model something genuinely contested and, left alone, it will hedge. It will tell you the evidence is mixed, that sources disagree, that the honest answer is "it depends." That hedging is not a bug. It's the model accurately representing the uncertainty in the world.

A knowledge skill can overwrite that. When you inject a confident factual claim into the model's context, you're not adding information to a blank slate — you're overriding the model's own calibrated uncertainty with borrowed authority. You've told it, in effect, "stop hedging, here's the answer." If the claim you injected is true, that's a gift: the model now knows your fiscal calendar, your process, your proprietary reality, and it stops waffling about things you've settled. If the claim is false, it's a disaster. The model will now assert your falsehood as flatly and fluently as it asserts that water is wet — and it will do it specifically on the topic you built the skill to help with.

That's the trap. Without the skill, the model would have told you the truth: "this is contested, be careful." With the bad skill, it tells you the false thing with total confidence. You didn't add knowledge. You added anti-knowledge — a claim that displaces the model's more honest default and leaves it worse informed than it started.

And here's the cruel part: the more authoritative and specific a knowledge skill looks, the more dangerous it is, because specificity and confidence are exactly what suppress the model's instinct to hedge. The impressive skills are the risky ones.

Let me show you one.

A worked example: the skill that tanks your reach

After a paid webinar, I was handed a polished skill for writing viral LinkedIn posts. It was well-built by every surface measure: clean structure, confident rules, sourced to a real report. One of its rules read, roughly: Links are back. Posts with three or more inline links get 441% higher reach. Include them inline; don't bury them in the comments.

Drop that skill in, ask for LinkedIn advice, and the model will now tell you — confidently, specifically — to stuff three or more external links into every post. So let's check the claim against the independent research, which is exactly what the skill's author didn't do.

It points the other way. Richard van der Blom's 2026 Algorithm Insights report, built on 1.3 million posts from 50,000 creators, found that a single external link in the body of a post reduces median reach by about 18.8% — meaning the no-link version wins. Broader analyses put the penalty higher, around 60% less reach for posts with external links. And in March 2026, LinkedIn's "Authenticity Update" specifically cracked down on external-link spam. The link-in-comments workaround the skill warned against? Also penalized now — but not because links suddenly help. Because all the off-platform link routes got tighter.

So where did "441% higher reach" come from? There is one analysis that found link-heavy posts outperforming: SayWhat's Q1 2026 study of roughly 400,000 posts. But read what its own authors say about why: the link-heavy posts were resource-rich posts — dense with genuinely useful, actionable material — and that quality is what earned the engagement. The links didn't cause the reach. The usefulness did, and the useful posts happened to contain links.

That single, hedged, correctly-caveated finding is what got turned into a hard rule. Watch the three things that happened to it on the way into the skill.

One: correlation became a directive. The research observed that good posts often contain links. The skill instructs you to add links to make posts good. That's the causation run backwards and frozen into a command. Adding links to a thin post doesn't borrow the magic of a resource-rich one; it just adds a reach penalty to a weak post.

Two: the caveats were stripped. The source was careful — it flagged the content-quality explanation, and it sat alongside a much larger dataset pointing the opposite direction. The skill kept none of that. It kept the exciting number and deleted the honesty around it. Nuance doesn't survive compression into a rule.

Three: the number is false precision. "441%" is a figure that doesn't trace cleanly to any of the underlying research, and it couldn't. You cannot extract a clean causal multiplier from observational data about which posts happened to do well — because you're selecting on the winners. The top posts share a hundred traits; picking one, links, and assigning it a precise percentage is numerology, not measurement. The decimals are there to feel rigorous, which is the opposite of being rigorous.

Put it together and you have a skill that takes a marketer, loads their AI assistant with a backwards, decontextualized, made-up-precise rule, and produces fluent, confident advice that plausibly lowers their reach. Ask that same model without the skill and it would have given the honest answer: link strategy on LinkedIn is contested and probably penalized, so lead with value and be cautious with outbound links. The skill didn't make the model smarter about LinkedIn. It made it dumber, and convincing.

(The same skill also insisted on posting six times a week. The 2026 research found optimal frequency had dropped to two-to-four. Once you see the pattern, you see it everywhere.)

How to spot a fail-dangerous skill

You don't need to be a domain expert to catch these. You need a short checklist, run on any skill before you trust it with your work.

Separate the claims from the choreography. Read the skill and mark every line that asserts something is true about the world — a number, a rule, a "this works / this doesn't." Those are your risk surface. The lines that just say "format it this way" or "use this tone" are safe; skip them.

For each claim, find the source — and count the sources. A knowledge skill living on a single report is fragile, especially in a fast-moving domain. One study is a data point, not a settled fact. If the skill cites nothing, that's your answer.

Ask whether a correlation got promoted to a directive. The tell is a rule of the form "successful things have X, so do X." Real causation is rare and hard-won; "X correlates with winners" is cheap and usually confounded. If the skill can't tell you why X causes the outcome, assume it doesn't.

Check whether the caveats survived. Go back to the source if you can. If the original hedged and the skill doesn't, someone deleted the truth to keep the punchline. Suspiciously clean confidence is a symptom, not a strength.

Run the mechanism smell test. Could the claimed cause actually produce the claimed effect? Some claims fail on contact with physics. A skill I saw insisted that lead magnets made with one AI tool outperform those made with another — a claim the platform's algorithm has no way to detect and therefore cannot reward. Mechanically impossible means fabricated, full stop.

Ask how fast it goes stale. Facts about an algorithm, a platform, a tool, or a market have a short shelf life. A skill that hard-codes this year's numbers is quietly wrong within a couple of quarters, and it will keep asserting the stale figures with the same confidence it had on day one.

The rule underneath all of this

Procedure skills need good taste. Knowledge skills need a verification discipline — and almost no one selling skills has one.

A knowledge skill is worth exactly as much as the truth of what it injects, and no more. Get the facts right and it's one of the highest-leverage things you can give a model: your real context, your real process, the things the model genuinely didn't know. Get them wrong and you've built anti-knowledge, and anti-knowledge is worse than an empty file, because it doesn't sit there doing nothing. It speaks, confidently, in your model's voice, on the one subject you trusted it with.

So before you trust a skill with facts, verify the facts. Or hand it to something that will.