The skills economy

Intelligence is getting cheaper. Capability is not.

A recent demo made the gap obvious. Someone gave Hermes, running DeepSeek, one long prompt for a Call of Duty-style first-person shooter. Thirty-two minutes and seven cents later, it returned a playable 3D game in one HTML file.

It is not a finished game. The frame rate struggles, the enemies are simple, and nobody should confuse it with a real studio release. That is not the point. The interesting part is how much specific behavior came out of a cheap model once it had a detailed procedure.

The author shared the prompt too. It specified the aesthetic, controls, weapons, map, HUD, enemy behavior, audio, performance constraints, and even the delivery format. That prompt is still a one-off instruction, not a reusable skill. But package the procedure with examples, scripts, assets, and checks, and the difference gets small fast.

This is what I mean by the skills economy.

Raw intelligence is becoming the commodity

For the past few years, most of the AI market has been organized around models. Which model is smartest? Which benchmark did it win? How large is its context window? How low is the token price?

Those questions still matter, but they matter less each month. Models keep getting more capable. Prices keep falling. More labs reach roughly the same frontier. Access to raw intelligence spreads quickly.

What remains scarce is knowing how to apply that intelligence to a real job.

A model can know how to code and still build the wrong thing. It can understand accounting and still produce the wrong workbook. It can know Swift and still miss the exact simulator, accessibility, signing, and review steps needed to ship an iOS feature.

A skill closes that gap. It gives an agent procedural knowledge: what to inspect, which tools to use, what order to work in, where people usually fail, and how to prove the result.

A prompt says what you want.

A skill teaches the agent how to get there again.

That distinction is why I think skills will become one of the main ways we sell intelligence and capability.

The early market

I wrote about this earlier when a few skill projects started escaping the small agent-builder corner of the internet.

Superpowers packages an opinionated software workflow: clarify the design, write a plan, use test-driven development, dispatch focused agents, review the work, and verify before claiming completion. It does not make the underlying model smarter. It makes the model work like someone who has learned a better process.

Compound Engineering packages a different operating system for software work. Its skills cover planning, execution, simplification, review, debugging, testing, shipping, and capturing what the team learned so the next run starts ahead.

Matt Pocock's grill-me is much narrower. It interviews you one question at a time until the hidden decisions in a plan become explicit. A small skill can change the quality of every task downstream because it fixes alignment before generation starts.

GBrain applies the same idea at a larger scale. It installs a set of skills that teach an agent how to ingest, retrieve, connect, and maintain personal or company knowledge. The database matters, but so do the procedures that tell the agent how to use it.

Hermes is another proof. Its skill library lets the same general agent switch from code review to email, document work, research, browser QA, media generation, or infrastructure operations without retraining the model. Each skill carries the hard-won details that separate a plausible answer from completed work.

There is now data behind the intuition. SkillsBench tested 7,308 agent runs across 84 tasks. Human-curated skills raised pass rates by 16.2 percentage points on average. Claude Haiku 4.5 with skills scored 27.7%, beating Opus 4.5 without skills at 22.0%.

The more important result was what failed. Skills generated by the model itself were flat or worse on average. The valuable part was not extra text. It was procedural knowledge curated by someone who knew the work.

The expertise is still there. Now it fits in a folder.

Skills are the ball bearings

The robotics boom has an obvious surface: humanoids, autonomous factories, drones, warehouses, and machines doing work that used to require people.

Under that surface sits a less glamorous layer. Bearings, actuators, gearboxes, sensors, and other parts have to work in almost every machine. Nobody posts a cinematic demo of a ball bearing. But the whole system depends on thousands of small components turning force into reliable motion.

Skills are starting to look like the ball bearings of the agent economy.

The model gets the attention. The skill quietly makes it useful.

A good skill can look underwhelming on disk. It may be a Markdown file, a few scripts, some templates, and a test. But it can encode years of judgment. It can tell an agent which shortcut is safe, which failure is common, and what evidence counts as done. Run it across thousands of tasks and that small package becomes leverage.

This is why I think the long-term value may sit one layer below the models everyone is watching. Models supply broad intelligence. Skills turn it into repeatable motion.

What the market will need

A folder of Markdown is enough to prove the idea. It is not enough to support an economy.

For skills to become a real market, they have to travel. A skill worth buying should work across Hermes, Claude Code, Codex, and whatever comes next. The open Agent Skills format is a good start.

Buyers will also need proof. A skill should show that it improves outcomes for a specific model, harness, and task. Stars and install counts do not tell you that.

Trust may be harder. Skills can run scripts, call tools, and touch private data. Provenance, permission scopes, signing, and security review will matter as much as output quality.

Composition and pricing come after that. Useful capability will often come from a chain of small skills with clean handoffs. Some will stay open source. Others will sell as subscriptions, per-run licenses, support contracts, or a share of the value they produce.

The winner may not be the person who writes the longest skill. SkillsBench found that focused packages with two or three modules beat broad documentation. The valuable author will be the one who can compress real experience into the smallest reliable procedure.

That creates a new kind of builder.

A great recruiter can package how to source and screen a rare role. A great designer can package how to critique a landing page. A security engineer can package an audit that catches the classes of bugs they have seen for twenty years. A filmmaker can package a visual language. A lawyer can package the steps and checks behind one narrow kind of filing.

They do not have to train a model. They teach an existing one how they work.

Capability becomes installable

Software turned business logic into something we could copy and sell. Skills do the same for parts of human judgment and procedure.

That does not make every skill good. Most will be thin prompt wrappers. Some will make the agent worse. Some will be unsafe. The market will have to learn the difference between a clever demo and a reliable capability.

But the direction is clear.

Today, two people can buy access to the same model and get wildly different outcomes. The difference is their context, tools, taste, and process. Skills are how those advantages become portable. They can be installed, tested, improved, shared, and eventually sold.

The next app store may not look like apps. It may look like folders full of Markdown, scripts, examples, and verifiers.

That sounds small next to a frontier model.

So does a ball bearing next to a robot.