There are millions of skills in the public domain now. People prompt a chatbot to write a SKILL.md, save it to a repo, and share it. It feels productive. It looks like expertise.
Most of it isn’t.
Ask the hard question: could a tax-preparer agent, running one of these public skills, actually prepare a client’s return well enough that a human could sign off with minimal scrutiny? For the vast majority of shared skills, the honest answer is no. They’re shallow. They capture the shape of a task, not the substance of doing it right.
Shallow vs. deep
A shallow skill says: “To prepare a tax return, gather the client’s documents, identify deductions, and fill out the forms.” That’s a description of the work, not the doing of it. An agent running it produces confident, generic, subtly wrong output.
A deep skill encodes what an expert actually carries in their head:
- The decision tree — if the entity is structured this way, these forms apply, which trigger these schedules.
- The exceptions — the edge case that shows up once a year and burns anyone who hasn’t seen it.
- The judgment — when the clean answer is wrong, and how an expert can tell.
- The substantiation — not just the answer, but how to defend it.
That depth is the entire value. It’s also the part a chatbot can’t generate for you, because it lives in your experience, not in the model’s training data.
Structure is what makes depth usable
Depth alone isn’t enough — a giant wall of expert text is unusable by an agent. Deep skills have to be structured: broken into sections with clear “when to use” signals, so an agent can retrieve the right part for the task instead of swallowing the whole thing. Progressive disclosure, not a data dump. This is the craft Skuilds is built to make easy: turning tacit, branching expertise into something an agent can navigate.
Living, not frozen
The last difference is maintenance. A real skill is never finished. Tax law changes. Best practices shift. The edge cases evolve. A skill saved as a static file the day it was written starts rotting immediately — and an agent running a rotting skill is worse than no skill, because it’s confidently out of date.
A living skill is versioned and maintained. You update it as your field moves, every change tracked, the endpoint always serving current guidance. Your expertise stays alive and keeps compounding in value instead of decaying in a drive.
Shallow, frozen text files are why people don’t yet trust skills for serious work. Deep, structured, living skills are how that changes. That’s the bar — and it’s the bar worth building to.