Today an AI agent integrating SourceryKit must read our docs top to bottom before it can act. We should ship SourceryKit as agent skills instead — small, self-contained instruction units (a SKILL.md: name/description frontmatter + body) that a coding agent loads automatically the moment a task matches. The agent gets the right guidance at the right time, scoped to what it's doing, instead of carrying the whole manual.
Skills are the emerging cross-ecosystem standard for this: the same SKILL.md format is read by Claude Code, OpenAI Codex, GitHub Copilot, and Gemini CLI (each just looks in a different directory). Shipping them makes SourceryKit first-class inside agent tooling, not something an agent has to go read about.
What to ship
A small set of skills scoped to how the SDK is actually used:
- Full-integration skill — the whole flow:
bootstrap_system → insert_trusted_endpoint → async_intercept_context → structured output (SourceryKitAgentResponse) → build_handoff_payload → evaluate_handoff. For a single-agent app or the orchestrator of a larger system.
- Role-scoped skills — one per stage (setup, recording HTTP calls, verifying claims). These serve multi-agent systems, where subagents own different stages: a verifier that only calls
evaluate_handoff should load only that guidance, not the whole pipeline. The handoff payload passed between stages is self-describing, so a downstream skill needn't re-teach the upstream steps.
Guiding principle
Skills should encode our own opinionated view of the correct way to integrate — grounded in, and cross-referencing, the cookbooks and docs; not a blind copy, not a bare link.
Out of scope
How skills are packaged, installed, or kept in sync per tool — decided by whoever picks this up.
Today an AI agent integrating SourceryKit must read our docs top to bottom before it can act. We should ship SourceryKit as agent skills instead — small, self-contained instruction units (a
SKILL.md:name/descriptionfrontmatter + body) that a coding agent loads automatically the moment a task matches. The agent gets the right guidance at the right time, scoped to what it's doing, instead of carrying the whole manual.Skills are the emerging cross-ecosystem standard for this: the same
SKILL.mdformat is read by Claude Code, OpenAI Codex, GitHub Copilot, and Gemini CLI (each just looks in a different directory). Shipping them makes SourceryKit first-class inside agent tooling, not something an agent has to go read about.What to ship
A small set of skills scoped to how the SDK is actually used:
bootstrap_system→insert_trusted_endpoint→async_intercept_context→ structured output (SourceryKitAgentResponse) →build_handoff_payload→evaluate_handoff. For a single-agent app or the orchestrator of a larger system.evaluate_handoffshould load only that guidance, not the whole pipeline. The handoff payload passed between stages is self-describing, so a downstream skill needn't re-teach the upstream steps.Guiding principle
Skills should encode our own opinionated view of the correct way to integrate — grounded in, and cross-referencing, the cookbooks and docs; not a blind copy, not a bare link.
Out of scope
How skills are packaged, installed, or kept in sync per tool — decided by whoever picks this up.