<aside> 🧩
Document purpose: Let ISAC grow by learning reusable ways to accomplish goals, not by collecting brittle macros.
Document status: Architecture/design specification. Current implementation status is controlled by the roadmap and stage evidence.
</aside>
A skill should contain a goal, preconditions, inputs, required tools, permissions, knowledge, procedure, decision points, failure handling, recovery, success criteria and historical performance.
flowchart LR
U["User teaches"] --> S["Skill system"]
O["ISAC observes repeated workflow"] --> S
R["ISAC recommends relevant skill"] --> S
ISAC may learn simple, safe procedures automatically when GCF has sufficient evidence. Learning a new procedure does not grant new permissions. Any capability expansion still requires governance and user approval where appropriate.
When a repeated workflow becomes stable enough, ISAC can say that it believes the workflow can become a reusable skill and offer to create/test it.
Skills should be versioned. New variants are evaluated against the current version and promoted only when outcomes improve. Failed variants are rejected or rolled back.
ISAC may recommend skills based on observed applications, repeated tasks, goals and user profile, but should explain what the skill would access before installation.
Imported skills start untrusted. Inspect manifests, permissions and network/file/process access, then sandbox/test before user approval.
GCF evaluates outcomes, confidence, regression and governance. ISAC should never silently turn a learned workflow into unrestricted automation.