Skills have become a common primitive in current agent tooling. Their folders now carry many different responsibilities.
Skills/
├── procedure
├── domain knowledge
├── personal preference
├── specialized agent
├── project state
├── output template
├── runtime code
└── finished result
When all of these fit under the word “Skill,” humans lose the vocabulary needed to design and maintain the environment around them.
In the Workplace-first Proposal, the workplace becomes first-class. A Skill remains a valuable provider-facing activation surface inside it.
Activation is not ownership
In this model, a Skill is a provider-readable activation surface. It helps a provider discover when and how to use an available capability.
The capability may be larger than that surface. It can include a reusable method, limits, controls, runtime support and several interfaces. The work it touches can belong to a Room or Site. Each execution produces a candidate; its results can remain ephemeral, become Material through a human transition or be delivered somewhere else.
The Skill activates the capability. It does not become the owner of all those responsibilities.
It does not own the human's intent either. The human chooses the outcome; the Workplace supplies context and boundaries; the Skill exposes one way to act.
The Workplace-first decomposition
The Open Workplace model separates responsibilities before an implementation generates provider files:
shared rule → Home
human belonging → Member
personal context → Desk
domain truth → Room Material
temporary state → Meeting
reusable method → Equipment
product truth → Site
external access → Route
activation → Skill projection
In Endroit, one Equipment package may project a Codex Skill, a Claude Command, a CLI route and a human-readable manual. Those interfaces expose one capability without becoming its canonical source.
Equipment
├── purpose
├── manual
├── controls
├── limits
├── optional engine
└── projections
├── Codex Skill
└── Claude Command
Why the boundary matters
If a Skill contains a Room’s decisions, removing or upgrading the Skill can silently move domain truth. If it contains personal preferences, sharing it can turn one person’s style into a Home rule. If it contains live product state, the projection becomes stale when the Site changes.
Explicit ownership keeps each change local:
- update the Room without repackaging the method;
- replace Equipment without losing its outputs;
- generate another provider projection without migrating the canon;
- let different agents activate the same capability through provider-specific surfaces.
An HACP Card answers a separate question: which semantic move did the human invoke? A Card may use a provider Skill or command as its surface without making that projection the owner of the capability or its work.
Skills still matter
Skills remain excellent activation surfaces. AGENTS.md and CLAUDE.md remain
useful front doors. MCP remains a useful tool boundary.
Workplace-first keeps a delivery mechanism from becoming the entire ontology of human-agent work.
The same boundary applies above the Skill. A provider or harness can execute the capability without becoming the owner of its context, output or continuity. Equipment owns the reusable method, never the result produced with it.
That separation lets a useful method compound through refinement without silently accumulating the candidates or decisions it helped produce.
Instructions guide the agent. Skills expose capabilities. The workplace keeps the owners of methods, domains, results and external truth distinct.
Once those responsibilities are clear, a Skill can focus on exposing a capability to its provider.
For a dated account of this separation in practice, see I Gave My Agents a Workplace.