The Only AI Skill Worth Paying For Is Context

The largest skill directory on the internet advertises 2,154,976 skills. By the time you read this the number will be higher, and it doesn't matter, because the same directory tells you in its own footer that it "does not certify that every skill is safe or high quality." It's indexing public GitHub repos. The number is inventory, not value.

And the sheer size is the argument, not a footnote to it. You don't accumulate two million of anything that's scarce or hard to make. A number that large means two things at once. First, these skills are trivially cheap to produce — which is precisely what "commodity" means. Second, they're piling up far faster than anyone is sorting the good from the worthless: the biggest directories apply no quality filter beyond "the repo has at least two stars." A market that floods with millions of uninspected skills isn't a rich market. It's a market without enough discerning buyers to punish the junk — which is exactly why the junk keeps getting made.

So the useful question isn't "which skill should I download." It's "which kind of skill is worth paying for." And once you sort them properly, the answer is narrow and specific — and it's probably not the kind anyone's trying to sell you.

One question sorts all of them

In an earlier piece I argued that the value of a skill is never the persona painted on top of it — it's the substance underneath: the knowledge, the procedure, or the context it actually carries. That's true, but it's not the whole story, because those three kinds of substance are worth wildly different amounts of money.

There's a single question that ranks them: who else can produce this exact skill?

The more people who can make the same skill you're being sold, the less it's worth. Run each category through that question and the pricing falls out on its own.

Procedure skills: anyone can make them

A procedure skill encodes how to do something — the steps, the format, the sequence. "Sort this brain dump into three priorities." "Structure a code review this way." "Draft the email in this shape."

Who else can make it? Everyone. The procedure is generic by definition; if it weren't, it wouldn't fit a downloadable template. Worse, the model usually already knows the procedure, or will do it correctly the moment you describe it in a sentence. You're paying for a formalization of something that was one instruction away from free.

This is where most of the two million live. It's the category being bundled into "50 skills that replace your marketing team" packs and sold off the back of a webinar — a cold-email writer, a meeting-notes-to-action-items skill, a "turn my rough notes into a polished post" skill. Occasionally handy. Almost never worth money, because the moment one is, someone posts an equivalent for free — and the model was going to do the job the second you asked it to. Procedure skills are commodities. Price: roughly zero.

Knowledge skills: perishable inventory

A knowledge skill injects what's true — facts, figures, rules, the current state of some domain. These feel more valuable, and briefly they are. But they carry two problems, and the second one is the one nobody mentions.

The first is accuracy. A knowledge skill is only worth as much as the truth of what it injects, and if it's wrong it doesn't just fail — it overrides the model's own honest uncertainty with confident nonsense. (That's its own whole failure mode, and worth understanding before you trust any skill that asserts facts.)

The second problem is decay, and it's structural. A knowledge skill that teaches the model something publicly knowable — a framework, a best practice, an algorithm's observable behavior — is living on borrowed time, because the next model absorbs that same public knowledge into its weights. You are selling the model something it is about to know for free. Every training run quietly eats a slice of your catalog. The skill you sold in January is redundant by the next release, still confidently reciting last year's version of a fact the model now knows better.

So knowledge skills are perishable inventory. Real value, short shelf life, and a fact-check required before every use. Price: a little, briefly, with your eyes open.

Context skills: the one durable asset

Now run the question on the third kind. A context skill encodes your particular reality — your chart of accounts, your customers, your process, your decision rules, your terminology, your private definition of what "good" looks like.

Who else can make it? Nobody. Not because it's technically hard, but because no one else has your context. And here's the part that makes it durable where knowledge skills are perishable: your context will never be absorbed into a model's weights, because it will never be in a training set. It's private. No release will ever eat it. The model can learn every public framework in the world and still not know that your team treats a signed order differently from a verbal one, or that "the Q3 problem" means a specific thing to six specific people, or that your reconciliation has one exception everyone knows and no one has written down.

That's the only category that's reliably worth real money. Not because context skills are fancy — most are mundane — but because they're yours, they don't decay, and the next model won't quietly make them free.

Everything about the economics points the same direction. Procedure: commodity. Knowledge: perishable. Context: durable. The value climbs exactly as the pool of people who could make the skill shrinks — and for context, that pool is one.

Why context is also the hard part (and therefore the real work)

If context is the valuable category, there's an obvious question: why isn't everyone selling context skills? Because you can't sell someone their own context. You have to extract it from them, and that turns out to be the actual skill.

A few months ago I spent several sessions interviewing an operator to build a set of skills around how she actually works. What I learned is that the value was never in the tool that generated the SKILL.md file — that part is nearly automated now, and commoditizing fast. The value was in two things the tool can't do.

The first is elicitation: knowing which questions surface the context that actually changes the model's behavior. People bury their most valuable decision rules inside throwaway phrases — "oh, I just always check X first" — and breeze right past them, because to them it's obvious, and obvious things don't feel worth saying. The load-bearing context is almost always the part the person doesn't think to mention. Getting it out is an interviewing skill, not a software feature.

The second is curation: knowing which of the extracted context to keep. Most of what a person tells you about their work is texture, not signal — it flavors the output but doesn't change it. A small fraction genuinely alters what the model produces. Keeping only that fraction is the difference between a skill that works and a bloated document that doesn't. Telling them apart is judgment — the same judgment that lets you look at a purchased skill and see whether it's carrying substance or a costume.

Elicitation and curation are where the real work lives, and neither one is on any directory, because neither one is a file. They're a process.

And it's a process no one at the frontier is going to run for you. The labs are making the engine better at everything in general, and they'll keep shipping memory features and personalization hooks along the way — but a hook isn't the work. "Better in general" and "knows your world" are different axes, and no amount of the first produces the second. A frontier lab is never going to sit down and interview you about the one exception in your reconciliation, the way your team actually scores a lead, or the word that means something specific to six people in your company — and then decide which of those genuinely change what the model should do. That's not a gap the next model closes. It's a permanent division of labor: the frontier keeps making the model smarter, and it still doesn't know you. Someone has to close that last gap by hand — turn your reality into something the model can use — and that work doesn't get automated away by a better base model. It's the rare job in this space that gets more valuable as the models improve, not less.

The build-vs-buy verdict

Put it together and the decision rule is short.

Procedure skills: don't buy. Write the sentence yourself, or let the model do it. You already own this.

Knowledge skills: buy rarely, and verify every time. Treat them as perishable, fact-check what they assert, and expect them to expire. Never trust one on a topic where being confidently wrong is expensive.

Context skills: build them — or have them built. This is the only category where paying makes sense, and what you're paying for isn't a file. It's the elicitation and curation that turn your reality into something the model can use. That's the work that doesn't commoditize and doesn't decay.

The two million downloadable skills are, overwhelmingly, the two cheap categories — the ones that are free, perishable, or both. The valuable one was never in the directory. It was in your head the whole time. The only question is whether you can get it out — and that's the one thing Solaris exists to do.