Skills
person concept tool org talk claim — click a node to jump to its page; hover an arrow for the relation
Skills are reusable, structured capabilities or procedures provided to or discovered for AI agents that enable them to perform specific tasks within an Agent Workspace or engineering pipeline. Across practitioners, skills serve as the primary unit of agent capability — analogous to SOPs or tools — and are distinguished from raw instructions by being encoded, retrievable, and composable.
Discovery vs. Design
A central philosophical position, articulated by Alex Damis, is that skills should not be designed upfront but discovered through doing the work. "I don't build skills, I discover them" — the claim being that the standard operating procedure emerges organically from iterative practice rather than top-down specification. The supporting infrastructure for this approach is Skill Lab, described as "really how I set up the scaffolding to capture the magic in the bottle as I'm discovering the magic". A concrete example of a skill produced through this process is the Alex Decision Lens, "a skill called Alex decision lens", which guides the agent through a handcrafted analytical procedure.
Skills as Persistent Encoded Knowledge
Boris Cherny advocates for writing reusable knowledge into skills or CLAUDE.md rather than simply re-instructing the model at runtime: "I tell it to write it to the quadmd or to like make a skill or something to do it differently". The underlying rationale is that this encoding is what enables agents to run reliably and indefinitely — corrections become durable rather than ephemeral.
Scope: Focused and Deep, Not Broad
Ryan Lopopolo at Harness Engineering advocates strongly for a small, focused skill surface: "we centralize our leverage around five to 10 skills. we don't go super wide on skills preferring to make the existing skills better". The reasoning is pragmatic — tooling changes too fast to maintain a wide skill surface area, so it is better to iteratively deepen a small core set. In practice, Harness gives each agent ticket "a couple of skills that enable it to manipulate our app", keeping the skill footprint minimal per task.
Skills and Data Retrieval
In Jason Lopatecki's self-improving agent architecture at Arize, skills do active data work: they "pull little temp files down into the repo", retrieving and structuring Traces and spans|traces and spans into files the agent can read. Lopatecki's position is that simply pointing an agent at raw observability data is insufficient — "you've got to kind of design the skill surface area in a way that Claude can really really work well". Skills mediate between the messy data environment and the agent's reasoning, and the context an agent operates on "is really driven by skills".
Points of Disagreement
| Question | Position | Speaker |
|---|---|---|
| How should skills come into existence? | Discovered emergently through iterative work; SOPs cannot be designed upfront | Alex Damis |
| How should skills come into existence? | Explicitly written or instructed by the user as corrections accumulate | Boris Cherny |
| How many skills should an agent have? | Small focused set (5–10), deepened over time | Ryan Lopopolo |
| How many skills should an agent have? | Implicit: skills grow as discovery proceeds; breadth is acceptable if grounded in real work | Alex Damis |