Patching Gemini's blind spots about its own API

What it does

The repo ships three “skills”—structured context bundles consumed by coding agents—to ground LLM output in current Gemini API patterns. One covers general API development, another targets the Live API’s WebSocket streaming stack, and the third maps the Interactions API across Python and TypeScript. The idea is to compensate for the fact that model weights are frozen at training time while SDKs and best practices keep shifting.

The interesting bit

Google bundles this as an unofficial side project, yet cites internal evaluations showing the skill boosted correct API code generation to 87% with Gemini 3 Flash and 96% with Gemini 3.1 Pro. It is essentially a curated, versioned cheat sheet that agents pull in at runtime to avoid thought-signature rot and deprecated model guardrails.

Key highlights

• Three focused skills: gemini-api-dev, gemini-live-api-dev, and gemini-interactions-api.
• Covers niche but sharp edges like voice activity detection, native audio features, and WebSocket session management.
• Bundles both Python and TypeScript SDK guidance for the Interactions API.
• Ships alongside a public MCP server (gemini-api-docs-mcp.dev) for live doc retrieval, though the skills work standalone.
• Explicitly not an officially supported Google product.

Caveats

• Skills are installed via third-party CLIs (Vercel skills or Context7), not as a direct library import.
• The vertex-ai-api-dev skill has already moved to a different repository.
• Google labels this as an unsupported side project.

Verdict Worth a look if you are building agentic workflows against the Gemini API and want to reduce hallucinated endpoints. Skip it if you are not using a skills-compatible CLI or MCP toolchain.