Connect Claude to Google Drive and check the sources behind its answer

Example result.

Opened: Depot refit scope note (14 Sep); Costs v3 (2 Oct); Decision record (30 Sep)
Found, not opened: Costs v2 (superseded by v3)
Could not read: Site survey pack – over the size limit, name only
Answer: the extension is priced as an option but not approved.

1. Connect the right account

In claude.ai, open Customize › Connectors, find Google Drive and select Connect. Sign in to the Google account you mean to use and grant the permissions. When it works, the button changes to Disconnect. Claude can reach only the files that Google account can open.

Then open the connector’s page under Your connectors and look at Tool permissions. Each group of tools can be Always allow, Needs approval or Blocked. Leave anything that creates files on Needs approval for now. Reading your Drive and adding files to it are different kinds of action, and the approval prompt is where you find out which one is about to happen.

In the chat, select +, then Connectors, and turn on Google Drive. The toggle only affects that conversation.

If your organisation hasn’t enabled the connector, ask the Owner. Don’t connect a personal account as a workaround.

Why this works: the permissions you set now decide what Claude can do later without asking.

What you should see: Google Drive under Your connectors, showing Disconnect, and the toggle on in your chat.

2. Pick known files and know the limits

Choose a question with a small, known source set, such as the unresolved decisions across a scope note, a cost sheet and a decision record. Open them yourself first and note which one sets the rules. “Latest” and “agreed” are not the same thing.

To attach specific files, select + › Add from Google Drive and pick them. Claude reads the current content each time you send a message. Know what arrives:

• Google Docs arrive as Markdown, without images.
• Sheets arrive as CSV, with every tab.
• Slides arrive as plain text.
• Above Google’s export cap of about 10 MB, Claude gets the file’s name and details but not its content.

Claude also can’t read comments or suggested edits, and a very large file may arrive incomplete.

Why this works: a big deck can be “attached” and unread, and nothing in a fluent paragraph will tell you.

What you should see: your named files attached to the message, and a note of any file large enough to be at risk.

3. Ask for a manifest before the answer

Use only these Drive files to answer [question]: [titles and links].
Do not create, change or share files.

Before answering, give me a manifest:
1. Every file you opened, with its last modified date.
2. Every file you found but did not open, with the reason.
3. Anything you could not read: too large, truncated or unsupported.

Then answer. For each material claim, quote a short passage and name
its file. Keep conflicting statements separate. Do not pick an
authoritative version from its date alone. Say what this set of files
cannot answer, and do not claim you searched all of my Drive.

[question] is what you need to know; [titles and links] are the files you chose in step 2.

Why this works: asking for the manifest first makes it an account of the work, not a justification written afterwards. Items 2 and 3 catch the quiet failures: a file found and skipped, or one that never arrived.

What you should see: three short lists, then an answer with a quoted passage for each claim.

4. Check the sources yourself

Open two or three files and find the quoted passages. Not “does it sound right”: find the sentence. Check that the nearby wording hasn’t been dropped, such as “provisional” or “subject to approval”.

Compare the manifest with your own list. If a file you expected is missing, or a file is listed under “could not read”, hold any conclusion that depends on it. Where decisions might live in comments, open the comments yourself.

Why this works: Claude’s list of what it read is useful, but it isn’t an access log. Opening the files is the check.

What you should see: each claim you’ll act on matched to a sentence you’ve read, and the gaps written down.

Check the result

• Every file you expected appears in one of the three lists.
• Each important claim has a quoted passage that you found in the file.
• Anything listed under “could not read” is treated as a gap, not as “nothing there”.
• Where two files conflict, both statements are shown, rather than the newest one winning.

The usual failure: Claude treats the most recently modified file as the decision. A proposal, a cost estimate or a draft with a newer date quietly overrules the agreed document. Spot it by checking the status words in the newest file: “provisional”, “draft”, “option” or “no approval yet”.

Your answer says [claim]. The named source says [passage and file].
Correct the claim to match the evidence, keeping its condition or
uncertainty. A newer date does not make a file authoritative. If
another file is needed, name the gap rather than filling it.

[claim] is the sentence you doubt; [passage and file] is what the source actually says, and where.

Fictional example. A manager at Harbrook Logistics asks what has been decided about the depot refit. Her three expected files say:

Scope note, 14 September:
"The refit remains within the existing storage area. Any expansion
requires a separate approved change."

Cost estimate v3, modified 2 October:
"Provisional estimate includes an optional extension. No approval yet."

Decision record, 30 September:
"Extension decision deferred pending a survey."

The manifest lists all three as opened, Costs v2 as found but not opened (superseded), and the site survey pack under “could not read”: it is over the export limit, so only its name arrived.

Claude’s first answer: “The refit now includes an extension, as set out in the latest cost estimate.” The repair prompt, quoting “No approval yet” from v3, produced: “An extension is priced as an option but has not been approved. The scope note requires a separate change and the decision record defers it pending a survey.”

What mattered:

1. Date isn’t authority. The cost estimate was newest, and its own words said it wasn’t a decision.
2. The unread survey stayed a gap. The answer says the survey hasn’t been read, not that it found no problems.
3. Comments were checked by hand. The connector can’t read comments, so the manager opened the scope note’s comment thread herself before sending anything.

Task kit for this job: Download (ZIP) · Reusable instructions

Go deeper

Sources

• Google Drive connector
• Get started with connectors
• What are projects?