Skill Creator — author and update workspace skills

Skill Creator builds the skill files an agent runtime loads: a SKILL.md with frontmatter (name, description) and an operational body, plus supporting files only when they materially improve reliability.

The workflow is structured. Clarify the capability, the likely trigger phrases, and the target path (~/workspace/skills/<name>/). Choose a narrow scope — one clear job per skill, split unrelated jobs into separate skills. Plan the layout before editing: only the operational core goes in SKILL.md; bulky docs, examples, and schemas go in references/; templates in assets/ only when the final output uses them; helper binaries or checked-in helpers in bin/ are preferred over prompt-side protocol instructions.

Body conventions depend on the skill type: tool-backed skills get Purpose, Tooling, Auth, Operating Rules; workflow-only skills get Purpose, Workflow, Output Contract, Operating Rules. Examples stay short and directly executable. Credential collection for a provider is handled by credentials.request_api_access, and once a connector is connected, a scaffold script generates a SKILL.md whose Tooling and Auth sections already carry the credential mechanics — those sections are left as generated.

The operating rules guard quality: preserve working commands and repo conventions, never invent binaries, paths, or auth flows, keep frontmatter minimal, and compile any Python CLIs you create with python3 -m py_compile before reporting success. A final sanity check confirms the description says what the skill does and when it triggers, and that the body tells the model what to do next rather than explaining the whole domain.