Feed Cursor Live Docs

Cursor's model learned to code on a training set with a cutoff date. Your framework doesn't check that date before shipping a new release. Ask it to build against a library that shipped a breaking change last month, and it will confidently generate syntax that was correct — six months ago. Cursor's docs feature exists to close exactly this gap.

Why the model guesses

Every LLM has a training cutoff: a point in time after which it has seen nothing. Ask it about a package released, or substantially changed, after that date, and it falls back on the closest thing it does know — an earlier version's API, a deprecated method name, a config format the library abandoned two majors ago. The output looks confident and plausible because the model isn't aware it's out of date; it just pattern-matches to what it learned.

This is invisible until it breaks something. A hallucinated parameter compiles fine in your head but fails at runtime, or worse, silently does the wrong thing. The fix isn't a smarter model — it's giving the model the current information it's missing, in context, at the moment you ask.

What the docs feature actually does

Cursor's docs feature lets you point it at any documentation site and have it crawl and index the pages, keeping the content available as context for future prompts. In practice this means Cursor can ingest hundreds of pages from a single doc site — real-world usage has shown it indexing several hundred pages from one crawl — building a working knowledge base of the actual, current API surface instead of relying on what was baked into training.

Once a doc source is added, you can tell Cursor to consult it explicitly: "follow the Supabase auth docs" pulls the indexed content into the conversation instead of the model's stale internal guess.

How to set it up

1. In the chat, type @ and select Docs.
2. Choose Add new doc.
3. Paste the URL of the documentation you want indexed — a database provider's setup guide, an auth SDK's reference, a framework's API docs.
4. Wait for Cursor to crawl and index the pages (larger doc sites take longer).
5. In your prompt, reference the doc explicitly: "using the docs I added, build the auth flow."

From then on, that doc source stays available to reference in that project, so you don't need to re-paste the URL every time you touch that library.

Where this pays off

• A library or SDK you've never used. Instead of the model inferring conventions from similar-sounding libraries, it works from the actual reference.
• A framework or API that shipped recently. Anything released or meaningfully changed after the model's training cutoff is a blind spot by default — feeding docs is the direct fix.
• A service with unusual or strict conventions (a specific auth provider's token flow, a database's particular migration syntax) where a plausible-but-wrong guess is expensive to debug.

Pitfalls

• Don't skip this step and then blame the model for a syntax error. If you're working against something released in the last few months, assume the model doesn't know it until you've fed it the docs.
• Adding a doc isn't automatic use. Cursor won't necessarily consult an indexed doc unless you tell it to — say "follow these docs" explicitly.
• Large doc sites take real time to index. Kick off the crawl before you need the answer, not mid-task when you're already blocked.