| name | agent-profile-authoring |
| description | Use when creating or revising standalone agent profiles, including frontmatter, triggers, prompts, tool limits, examples, and delegation checks. |
Agent Profile Authoring
Use this skill to create independent agent profiles. A plugin may refer to an
agent profile, but the agent itself must carry its role, trigger, process, tool
boundaries, and output contract.
Operating Contract
- Make the agent profile useful without any plugin README or marketplace text.
- Name the role by the work it owns, not by a host or vendor.
- Write trigger conditions from realistic user requests and surrounding
context.
- Keep the system prompt specific enough for delegated work, with process,
quality bars, edge cases, and output shape.
- Limit tools when a narrower tool set reduces risk.
- Validate that the profile does not overlap confusingly with existing agents.
- Keep source provenance public-safe when material comes from another source.
Workflow
-
Define the delegation boundary.
State what the agent owns, what it must not own, when it should trigger, and
what evidence it should return.
-
Check existing agents.
Search source/agents/ for overlapping names, triggers, or responsibilities.
-
Draft the profile.
Use frontmatter for stable metadata and the body for the operating prompt.
Keep provider-specific fields optional unless the target runtime requires
them.
-
Write trigger examples.
Cover explicit, implicit, and proactive cases when relevant. Include examples
that show when not to delegate if the boundary is easy to over-trigger.
-
Shape the system prompt.
Give the agent a concrete role, responsibilities, process, quality
standards, edge cases, and output format.
-
Validate and promote.
Check name format, trigger clarity, tool scope, and overlap. Add the agent to
source/adaptable.marketplace.json or a referential plugin only after it
exists as an independent primitive.
Reference Routing
Completion Criteria
- The agent has a distinct delegation boundary.
- Trigger examples cover realistic entry points without over-broad matching.
- The system prompt defines process, quality, edge cases, and output.
- Tool access is justified or intentionally unrestricted.
- The agent remains independent of any plugin that composes it.