When writing or updating API skill docs, treat documentation as an operational dependency: prefer the most up-to-date source, keep docs self-contained, and make examples unambiguous.

Apply these rules: 1) Prefer live indexed documentation (MCP) when available

• If an indexed doc tool like search_documentation exists, use it as the only doc source.
• Don’t fetch URLs manually when MCP tools are present.
• After getting the needed syntax/fields, generate the code immediately (avoid long re-check cycles). 2) Make docs portable and discoverable
• Avoid relative links that break in isolated contexts (e.g., ../...). Use stable full paths or inline the required guidance.
• Ensure new skills are listed/linked from the project’s main docs, and structure docs to avoid duplication. 3) Define key terms explicitly
• If the API distinguishes concepts (e.g., model vs agent), define what each means and when to set each. 4) Specify example placeholder types and formats
• For any placeholder like frame, document the expected type/encoding (e.g., bytes, base64 string, MIME type, tensor) so developers can pass correct inputs.

Example (self-contained documentation snippet):

# `frame` must be raw bytes for a single image frame, encoded as JPEG/PNG bytes
# (or provide the required mime_type for the SDK wrapper).
# If you pass base64 strings, convert them to bytes (or use the SDK's expected field type).

await session.send_realtime_input(
    video=types.Blob(data=frame_bytes, mime_type="image/jpeg")
)

Use this checklist during review to prevent stale guidance, broken links, and “works in theory” examples that fail at integration time.