Skip to main content

Skills

Read this if: you need the architecture boundary for reusable instruction bundles.

Skip this if: you are looking for runtime-enforced workflows or in-process extension code.

Go deeper: Playbooks, Gateway plugins, Sandbox and policy.

Skill boundary at a glance

SurfaceWhat it providesEnforced by runtime?Best use
SkillReusable instructions and workflow guidanceNoTeaching the agent how to approach a task
PlaybookDeterministic workflow specYesMulti-step controlled execution
PluginIn-process gateway extensionYes, at gateway boundaryNew routes/tools/commands

Purpose

Skills are instruction bundles the agent can load on demand to perform specialized workflows consistently. They are guidance, not enforcement. The safety boundary remains in tool policy, approvals, and sandboxing.

In the compiled runtime prompt, skills live in the system prompt as Skill guidance: sections. They are intentionally lower-precedence than:

  1. system/runtime safety rules
  2. tool schemas and tool contracts
  3. the current user request

That keeps skills useful for workflow consistency without letting them override policy or tool contracts.

Load sources and trust posture

Skills can come from several locations, with more specific locations winning on conflicts:

  1. bundled skills
  2. user skills directory
  3. workspace skills directory, but only when workspace trust is explicitly enabled

Workspace skills are the main risk boundary. They come from the current checkout and must be treated as untrusted until the workspace is deliberately marked trusted.

The bundled tyrum-docs skill points the agent at https://docs.tyrum.ai when the user asks how Tyrum is meant to be used. It is product guidance, not a policy override.

Key rules

  • Skills must never be treated as a bypass around tool policy or approvals.
  • Skills must be framed as advisory workflow help, not as hidden policy or hard requirements.
  • Operator surfaces should show skill provenance so users can distinguish bundled, user, and workspace sources.
  • Marketplace discovery is acceptable only with reviewable provenance and explicit enablement.
  • Skills must reference secret handles rather than embedding raw secret values.