Skip to main content
The capability spec is the declarative definition of a capability. It describes what organizational ability should exist, what outcome it must deliver, and the boundaries within which it may operate. It is analogous to source code, or to an infrastructure configuration file: it says what must be true without prescribing every runtime decision.

​
What the spec contains

  • The Job to be fulfilled
  • The desired organizational outcome
  • Beneficiaries
  • Business value
  • Triggers
  • Required inputs
  • Expected outputs
  • Preconditions
  • Completion conditions
  • Policies
  • Constraints
  • Out-of-scope behavior
  • Allowed systems
  • Prohibited actions
  • Escalation conditions
  • Capability owner
  • Human roles
  • AI roles
  • Reviewers
  • Approvers
  • Stakeholders
  • Skills
  • Tools
  • Integrations
  • Credentials required
  • Knowledge sources
  • Memory access
  • Success criteria
  • Acceptance criteria
  • Evidence requirements
  • Risk classification
  • Review requirements
The trust contract is the part that distinguishes a capability spec from a prompt. It states in advance what would count as proof that the capability worked.

​
The spec is versioned

A capability may have many spec versions, but only approved versions may be instantiated and promoted. Each version is a distinct, reviewable artifact — which is what makes “what did we intend, and when did that change?” an answerable question.

​
Spec and recipe are different

This distinction is load-bearing. The spec defines the desired organizational ability. The recipe defines the approved way of delivering it. Changing how the work gets done does not change what the capability is for.

​
In the product today

You author a capability spec by talking to Joyvis in jstm agent plan, or in the web builder. The conversation produces a structured spec; committing it with jstm agent plan done is what turns the draft into a real agent definition. jstm agent spec <agent-id> exports the committed definition, and jstm agent view <agent-id> --definition shows it inline.

​
Next

Agents and recipes

The governed intelligent realization of a spec.

Draft an agent

Author a spec in conversation.