Skip to content

Skills

Target

This page describes intended behaviour or structure. No implementation is claimed unless a section is explicitly marked Current implementation or Evidence.

A skill is reusable expertise: how something should be done. It is not an agent (a role and configuration) and not a tool (an executable capability with side effects). A skill grants no execution power; it may require tools, models and knowledge, which policy then grants or denies.

What a skill contains

Instructions, workflow guidance, examples, references, scripts and templates (declarative, executed only through governed tools), tool requirements, model requirements, knowledge requirements, compatibility metadata, tests and evaluations.

Taxonomy (target)

Area Examples
Engineering / .NET modernization, package upgrade, architecture review, performance investigation, testing
Azure function debugging, Application Insights investigation, cost analysis, deployment
Database SQL performance, migration, recovery
Business requirements analysis, estimation, proposal generation, project planning
Company ConnectSoft engineering standard, security, architecture

Skills are reusable between agents; an agent version binds a skill version range.

SkillOps lifecycle

create → develop → test → evaluate → package → sign → publish → deploy → monitor → update → deprecate. Published skill versions are immutable; they have owners, dependencies, compatibility metadata, evaluation results and trust metadata. Publication is a Control plane responsibility (registries); evaluation gates are in Trust & Operations.

Current implementation

No Factory skill exists. The AgentKit and OpenClaw repositories below are earlier, non-Factory experiments (user decision D19, 2026-10-09) and are reference input only.

  • AgentKit .ai/skills/: 84 skills; bindings.json binds 40 of them to 38 agents; 44 are unbound (often direct developer tools, which is not a defect by itself).
  • dist/connectsoft-ai-agentkit/skills/*/contract.json: 36 contracts at version 0.1.0 with reviewPending: true; documented: false on all 84; the four skills bound to connectsoft-tech-lead are among the unreviewed.
  • OpenClaw runtime pack holds 48 skills; openclaw-skill-diff.md records three same-id content divergences (api-library-generator, documentation-site-generator, library-generator).
  • ConnectSoft.AISkillsLibraryTemplate produces typed AIFunction tools, despite its name; its "generation" is simulated.
  • The agents framework's file-skill script runner throws NotSupportedException; scripts inside skills do not execute.
  • The three status vocabularies of those experiments are not reconciled; the Factory lifecycle is Draft → InReview → Published → Deprecated (registries, D19).