A skill that fetches current AI SDK/API documentation before answering any question about Anthropic, OpenAI, or Google APIs. Prevents stale training data from producing wrong model IDs, deprecated parameters, or missed APIs.
AI coding assistants have a training cutoff. Documentation changes faster than models retrain:
- Model IDs get renamed or deprecated (
gemini-1.5-flash→gemini-2.0-flash) - Parameters get removed (
budget_tokensdeprecated on Claude Opus 4.6) - New APIs replace old ones (OpenAI Responses API replaces Chat Completions for agents)
- Built-in tools go unnoticed (Anthropic's native
web_searchandweb_fetchtools)
A 30-second doc fetch before answering catches all of these. See CREATION-LOG.md for the real-world failure that motivated this skill.
| Provider | What gets fetched |
|---|---|
| Anthropic | Models, tool use, streaming, prompt caching, batch API, extended thinking |
| OpenAI | Models, Responses API vs Chat Completions, function calling, streaming, batch |
| Gemini models, Gemini API vs Vertex AI, function calling, structured output |
Detection is automatic — the skill infers which provider is in scope from imports, environment variable names, and model name patterns in your code or question.
cp -r live-docs-lookup ~/.claude/skills/Copy the live-docs-lookup/ directory into your tool's skills/plugins folder. Check your tool's documentation for the exact path.
The skill activates automatically when you ask about:
- Which model to use
- Tool use / function calling / structured output
- Streaming responses
- Prompt caching, batch processing, rate limits
- SDK setup, debugging, or planning for any AI API integration
- Anything that involves Anthropic, OpenAI, or Gemini
You can also force it at any time with the slash command:
/live-docs-lookup
Useful when you want a fresh doc check before starting a new feature, or when the skill didn't auto-trigger but you'd like it to.
If you use the obra/superpowers plugin system, you can install live-docs-lookup separately and it will work alongside it. It will automatically trigger before brainstorming, plan writing, or other SDK-related workflows — so plans are built on current docs.
The evals/ directory contains 9 test prompts (5 Anthropic, 2 OpenAI, 2 Google) used during development. Pass rate with skill: 100%. Without skill: 45%. See CREATION-LOG.md for details.