As a skill library grows from five entries to a hundred, the rate at which the right one actually gets used falls from 29.6% to 3.3%. That is not a claim about badly written prompts. It is a measured curve across 8,135 trial records, and the entries in the hundred-item pool are the same entries that worked in the five-item pool. Nothing about them changed. Finding them did.
I have argued before that a pile of prompts is best on the day you buy it and decays from there. That was the polite version. The research published this month says something harder: a pile does not just sit there failing to improve. Past a certain size it actively interferes with itself, and every entry you add is a small tax on every entry already in it.
What was actually measured
Demystifying Agent Skills: Why They Work—Until They Don’t is a 2026 preprint from a group spanning Princeton, UC San Diego, Stanford, USC and Johns Hopkins. It ran controlled experiments across benchmarks, agent harnesses and models, normalized 8,135 trial records, and set out to separate two things the industry habitually merges: whether a skill is any good, and whether it gets used.
The headline curve is the retrieval one. Five-skill pool: the intended skill is actually used 29.6% of the time. Hundred-skill pool: 3.3%. Same skills, same tasks — the only variable is how many neighbours each entry has.
Sit with the direction of that. Most of us have assumed a large library is at worst neutral: some dead weight, no harm done, nice to have it there when you need it. The measurement says the dead weight is not inert. It is noise in the retrieval channel, and the channel is the only route your library has to the work.
The bottleneck is not where the market points
This is the part worth internalizing, because it relocates the problem.
The entire prompt-and-skill economy is organized around entry quality and entry count. Better prompts. More prompts. Prompts written by someone with a track record. All of that operates on the object. None of it operates on the channel — and the study finds retrieval degrading as a bottleneck separate from skill quality. You can improve every entry in a hundred-item library and move this curve not at all.
There is a sharper corollary in the paper, easy to skim past: exact ground-truth invocation turns out to be neither sufficient nor necessary for downstream success. Retrieving the right skill does not guarantee a good outcome, and a good outcome does not prove the right skill was retrieved. Which means the metric most tooling reports — did it find the thing — and the metric you care about — was the work any good — are measuring different phenomena. If you have been reassured by one, you have not been told about the other.
What skills are actually for
The second finding is the one I would put on a wall.
When skills helped, the study attributes 65.7% of those cases to procedural anchoring — the skill turning a noisy, improvised trajectory into a stable, repeatable one. Explicit knowledge injection, the thing everyone believes they are buying, accounts for 4.5%.
That is a category correction, and it lands on something I have said imprecisely. I have described the core problem as amnesia — your AI forgetting what you told it. Close, but off by one turn. On this evidence the model is not mostly missing facts. It is missing a fixed way of doing the work. It does not forget what you know. It forgets how you work, and improvises something plausible in the gap.
Which reframes what a library is for. You are not accumulating knowledge the model lacks. You are trying to hold a procedure still. And a procedure only has to exist once to hold still — whereas a fact might plausibly benefit from a hundred variations. The whole case for volume rests on the 4.5% mechanism, not the 65.7% one.
Why the count was always the wrong axis
Put the two findings together and the marketing arithmetic inverts.
If skills work mainly by anchoring procedure, then what you need is the right procedure, reliably reached. If retrieval precision collapses with pool size, then reliably reached is exactly what a large pool cannot deliver. The count is not a proxy for value. Past a small number it is a proxy for interference — and a vault of thirty thousand prompts is not thirty thousand assets, it is a retrieval problem with a price tag.
None of that makes a modular library the wrong shape. It is the right shape; that is the whole design. It is the right shape with a discipline attached — an index, a retirement rule, a record of what actually fires. Without those, the honest description of a growing library is a system that gets worse the more you invest in it. That is an uncomfortable sentence to write as someone who ships capability packs, and it is the sentence the data supports.
What the study does not say
Before anyone runs off and deletes things, the limits — which are real, and which I would rather state than have pointed out.
This measures agent-skill retrieval in controlled experimental conditions, not your prompt library on a Tuesday. Retrieval precision is a mechanism metric, not a business one; the paper’s own finding that invocation is neither sufficient nor necessary for success cuts both ways, and it means “3.3% precision” is not the same as “3.3% as useful.” The 5-to-100 range is what was tested — nothing here identifies the exact size where your setup turns the corner, and the curve says nothing about libraries organized by strong human indexing rather than automated retrieval. And a preprint from August is a preprint from August.
What survives all of those caveats is the direction and the mechanism: retrieval degrades with pool size, independently of quality, and it is the channel every entry must pass through. That is enough to change how you build. It is not enough to justify a number in a sales page, which is why you will not find one in mine.
The move this actually implies
Not “own fewer prompts.” Know which ones fire.
An entry no retrieval path reaches is not a dormant asset waiting for its moment. It is a competitor to the entries that do fire, and it costs its neighbours precision on every request. That reframes library maintenance from tidying — optional, aesthetic, eternally deferred — into something closer to the audit nobody schedules: find the entries that actually get invoked, retire the ones that never do, and give what remains a real index instead of a folder.
And then the part that does not come out of any paper. A procedure that stays fixed because you wrote it down once is worth more than fifty variations of it, and the way you find out which procedures you actually have is to watch what you keep correcting. Corrections are the raw material; the procedure is what they compound into. Volume was never the asset. The procedure was — and a procedure nobody can retrieve is not an asset yet.
If you build with AI, this is an index-discipline problem in your tooling — the software-engineer track. If you run a business on AI, it is the same question pointed at your prompt library and your custom instructions — the solopreneur track.
Both roundups covered the paper the week it landed: the developer edition and the solopreneur edition.
A pile does not sit there. It grows, and it crowds out the few things in it that were working.
Frequently asked questions
Does a bigger prompt or skill library make my AI better? The evidence says the opposite past a small size. A 2026 multi-institution study normalizing 8,135 trial records found that as a skill pool grows from 5 entries to 100, the rate at which the right one is actually used falls from 29.6% to 3.3%. The entries are not getting worse. Retrieval is — and retrieval is what stands between an entry and the work. Adding a hundredth entry makes the other ninety-nine slightly harder to find.
If skills don’t mostly supply knowledge, what do they do? They stabilize procedure. The same study attributes 65.7% of cases where skills helped to procedural anchoring — turning a noisy, improvised trajectory into a repeatable one — against 4.5% for explicit knowledge injection. The practical reading: you are not buying facts the model lacks. You are buying a fixed way of doing the work, which means one well-retrieved procedure beats fifty entries nobody can find.
So should I delete most of my prompt library? Not blindly, and the study does not say that. It measures retrieval precision, not business value, and it says explicitly that invoking the exact intended skill is neither sufficient nor necessary for a good outcome. What it does support is auditing by use: an entry no retrieval path reaches is not an asset, and it is quietly making its neighbours harder to reach. Start by finding out which entries actually fire.