Skills and Capability Packages

Fact-check status

Reviewed claim by claim on 2026-09-02 against the primary or official sources identified inline as S1 and S2. Recommendations are design guidance unless a cited standard makes them normative.

Key jargon

Term	Plain-language meaning
Skill	A reusable instruction package for a bounded kind of work. S1
Capability	An action the system can actually perform through code, a tool, or an integration. S1
Manifest	Metadata declaring identity, version, dependencies, permissions, and entrypoints. S1
Dependency	Another component or service required for correct operation. S1

Key concepts

Instructions describe behavior; capabilities create effects. Keep those layers explicit. S1
Versioned packages make specialized behavior discoverable, reviewable, testable, and revocable. S1

Concept map

flowchart LR
    A["Discover skill manifest"] --> B["Check dependency and permission"]
    B --> C["Execute bounded capability"]
    C --> D["Record version and outcome"]

Diagram semantics verified against S1 and S2 on 2026-09-02.

A skill is a reusable capability package: scoped instructions, prerequisites, reference material, scripts or tools, templates, examples, and verification rules. S1

Good skill properties

Narrow trigger and explicit non-goals. S1
Minimal required context and permissions. S1
Versioned dependencies and authoritative references. S1
Reusable assets instead of repeated generation. S1
Deterministic checks around model judgment. S1
Safe failure and escalation behavior. S1
Test cases showing when the skill should and should not activate. S1

Skill versus prompt

A prompt is one instruction artifact. A skill defines a repeatable operating method and may include prompts, tools, data, and tests. S1

Exercise

Package a “meeting-summary” skill. Include accepted input, privacy rule, output schema, unsupported decisions, example, quality rubric, and deletion/retention behavior.