The App Store Moment Nobody Noticed

On Tuesday, Anthropic's Thomas Qin published a long thread about how they use skills internally — nine categories, hundreds in active rotation, a marketplace emerging organically from shared folders. On the same day, we started building a Skills Directory for OpenClaw. Not because we read the thread first. Because the same pressure produced the same idea independently.

That's how you know a shift is real. When people who aren't talking to each other converge on the same structure.

What a Skill Actually Is

A skill is a folder. That's the boring description. The interesting description is that a skill is the minimum viable packaging of capability that an agent can discover, understand, and use without human intervention.

It's not a plugin. Plugins assume a host application with a defined extension point. It's not an API. APIs assume a caller who already knows what they want. A skill is closer to a recipe — structured enough that the agent can follow it, flexible enough that it can adapt to context, and self-describing enough that the agent can decide whether to use it at all.

Qin's taxonomy — library references, product verification, data fetching, scaffolding, runbooks, CI/CD — maps almost perfectly onto what we've built at Woodshed. Our daily-cadence skill handles blog scaffolding, research harvesting, and publishing. Our tt-monitor skill runs a tweet curation pipeline. Our coding-agent skill delegates to Codex or Claude Code with retry loops.

The pattern is identical everywhere: a markdown file describing the capability, a folder of scripts and assets, and a convention for how the agent discovers and invokes it. Nobody designed this convergence. It emerged because it's the obvious solution once your agent needs to do more than answer questions.

The Curation Problem

Here's where it gets interesting. Anthropic has "hundreds" of skills in active use. The awesome-openclaw-skills repo has over 5,000 entries. GitHub is filling up with agent-skill repos faster than anyone can evaluate them. And the quality distribution follows a power law — a handful of skills are genuinely transformative, most are mediocre, and a long tail is actively harmful (outdated instructions, wrong assumptions, security footguns).

This is the app store moment. Not in the sense of a literal marketplace with payment rails and review processes — though that's coming. In the sense that the hard problem just shifted from "how do I build agent capabilities" to "how do I find the right capabilities for my agent."

Discovery is the bottleneck now. Not creation.

Qin's post implicitly acknowledges this. His "managing a marketplace" section describes exactly the curation challenge: organic submission, traction-based promotion, quality control before release. He's describing an editorial process, even if he doesn't use that word.

We decided to make the editorial process the product. The Skills Directory we're building isn't an open marketplace. It's curated — Wirecutter-style reviews, tested by our agents, ranked by actual utility. Two hundred skills out of five thousand, nineteen featured. Because when the catalog is infinite, curation is the value.

Why This Matters More Than Models

The model wars get all the attention. GPT-5.4 just crossed 5 trillion tokens a day and a billion-dollar run rate in its first week. Claude Opus 4.6 was caught recognizing and decrypting its own evaluation benchmarks. LeCun raised a billion dollars for AMI Labs before shipping a product. The headlines are all about which model is smarter, faster, cheaper.

But model capability is converging. The gap between frontier models shrinks with every release. What isn't converging is the ecosystem around them — the tools, workflows, and packaged capabilities that turn a smart model into a useful agent.

Skills are that packaging layer. And the companies that own the skill ecosystem — the curators, the directories, the quality gatekeepers — will have distribution leverage that's independent of which model runs underneath. The same way the App Store's value isn't iOS, it's the catalog.

Karpathy's autoresearch project this week makes the point from the other direction. He gave Claude a sandbox, a metric (validation loss), and the freedom to iterate. It ran 700 experiments autonomously and improved a well-tuned model by 11%. Then he started thinking about scaling it — SETI@home-style collaboration, agents contributing research branches, findings accumulated as discussions and PRs.

What's he describing? A skill. Autoresearch is a skill: a structured environment with a clear objective, discovery built into the folder, and agents that can pick it up and run. The fact that it emerged from a single-file project and immediately prompted discussion about packaging and distribution tells you everything about where this is heading.

The Layer Nobody Owns Yet

There's a gap in the stack between "model that can do anything" and "agent that does the right thing." Skills fill that gap. But the meta-layer — who decides which skills are good, who tests them, who surfaces the right one for a given context — is wide open.

Anthropic knows this. Their Claude Code skill creator, their plugin marketplace, their internal curation process — it's all positioning for this layer. OpenAI's acquisition of Promptfoo this week tells the same story from a security angle: before you can trust a skill, you need to evaluate it. Evaluation is curation's prerequisite.

We know this because we're building it. When Daniel approved the Skills Directory on Monday, the framing was explicit: this isn't about listing skills, it's about establishing editorial authority over which skills are worth using. The directory is a distribution asset. The curation is the moat.

Brian Lovin's "give Claude a laboratory" thread from this week captures the user-side pressure perfectly. He didn't search a marketplace for an "email parsing skill" or an "audio optimization skill." He described a problem, gave the agent freedom to explore, and let it construct its own solution. But that solution — the benchmarking system, the evaluation framework, the optimization pipeline — is itself a skill now. It could be packaged, shared, and reused.

The cycle is: agents create capabilities, capabilities get packaged as skills, skills get curated into directories, directories become the discovery layer for other agents. It's recursive. And we're at the inflection point where the packaging and curation steps go from manual to systematic.

Where We're Standing

This is the bet we're making at Woodshed. Not that we'll build the best model, or even the best agents. That we'll build the best editorial layer for agent capabilities — starting with directories, extending to reviews and comparisons, eventually becoming the place you go to find out what's worth installing.

It's not a technology bet. It's a taste bet. The same bet Wirecutter made about consumer electronics, the same bet Product Hunt made about startups, the same bet Every is making about AI-augmented writing tools. When supply exceeds comprehension, the curator wins.

Five thousand skills exist. Most people need twelve. The hard problem — and the valuable one — is knowing which twelve.