Connecting Your Knowledge: Sources and Connectors
Uploaded knowledge is a snapshot; a connector is a window — and every official connector defaults to the same boundary: read broadly, act narrowly, never send on its own.
Two lessons back, the scoping question was upload versus paste versus connector, and connectors got a promissory "we'll cover this properly later." This is later. Once you understand what a connector actually does — and, just as important, what it deliberately can't do — it becomes the natural answer for every source in your work that changes without you touching it.
Five Sources, Two Freshness Guarantees
Everything Claude draws an answer from falls into a small number of source types, and they split cleanly into two categories: frozen at a point in time, or live right now.
Uploaded files and Project knowledge are both snapshots — accurate the moment you add them, and only as accurate as your last refresh after that. Connectors and web search are the opposite: they read the current state of the source every time you ask, which is exactly what you need for anything that changes on its own. The practical rule that falls out of this is the one from the earlier knowledge-scoping lesson, extended: if the answer needs to reflect something that changed an hour ago, no amount of careful uploading solves that. You need a connector, or a fresh search, full stop.
Google Drive and Gmail are the two most immediately useful connectors for most power users, because they sit on top of material you're already generating constantly — documents you're not going to duplicate into a Project's knowledge base, and an inbox that changes by the hour. Connecting them means Claude can find, read, and cross-reference that material live, inside a chat, instead of you copy-pasting it in every time.
The Boundary That Makes This Safe to Use
The reason connectors are worth granting access to — instead of a reasonable instinct to keep everything walled off — is that every official connector ships with the same conservative default: read broadly, act narrowly, and never send or publish anything on its own.
The Drive connector can find, read, summarize, and cross-reference files across your whole Drive — genuinely powerful for research and synthesis. What it cannot do is create, edit, move, rename, or delete a single file. It's read-only by design, which means granting it access can't accidentally result in a reorganized or overwritten file structure.
The Gmail connector goes one step further, into drafting — it can read your inbox and prepare a reply for you. But "prepare" is the operative word: it never sends. Every draft sits there until you personally review and send it. That boundary is what makes it safe to connect an inbox that has real consequences attached to what leaves it — the worst-case outcome of a bad draft is a message you delete before it goes anywhere, not a message that goes out on its own.
Deciding What to Actually Connect
Having access to a connector isn't the same as needing it, and the discipline from the knowledge-scoping lesson applies here too: connect what a real task in your workflow needs, not everything that's technically connectable. A calendar connector is genuinely useful if you're constantly asking "what does my week look like" — it's dead weight if you never ask that question inside a Claude chat in the first place. And some sources should stay disconnected on purpose regardless of convenience: anything sensitive enough that the cost of a mistake outweighs the time it would save, with no real task pulling you toward connecting it anyway.
The habit worth building is the same five-second pause from the scoping lesson, aimed at a slightly different question: before granting a connector, ask what specific task it unlocks, and whether that task is one you actually do. "It might be useful someday" is how an account ends up with six granted connectors and three of them unused — each one a permission boundary you granted for a task that never materialized.
With research, writing, analysis, and now live knowledge sources all covered, the pieces are in place to answer a question that comes up constantly for anyone using more than one AI tool: when does Claude actually win, and when does something else?