Loading post…
We put three team handbooks in Google Drive and the same three in Warmhand packs, connected each to Claude, and asked 58 questions through the same model. On the 43 real questions, both sides answered every one correctly, and Drive never opened a document from the wrong handbook.
The differences were in how each side got there. Drive handed the model a median of about 14,000 tokens per answer; the pack about 2,300. The pack lost on ranking: in a real account that holds other packs, its top search result was usually from an unrelated one. And when we asked about things the sales handbook does not cover, the two sides told the person so at different moments.
This is the third post in our head-to-head series. The second put the same sales handbook up against a Claude Project.
For most teams, Drive is where the handbook already lives. Claude's Google Drive connector searches every file the person can open and reads what it finds, so there is nothing to set up beyond connecting it. In our research post we expected the weak spot to be scope: with no way to point the connector at one folder, the AI might answer from the wrong document. This test measures that.
Three handbooks, on both sides.
The company and support handbooks are there as distractors. Several of them sit close to sales topics: billing, on-call, working with sales.
The two sides.
Held equal.
The questions.
Scoring. The real questions were scored as in post 2: the key string from the answer passage, then a fixed list of key facts, then a read-through of every miss on both sides. For every chat we also pulled the full transcript, with each tool call and what it returned, to see which files were opened and how much text reached the model.
| Google Drive | Warmhand | |
|---|---|---|
| Correct on the 43 real questions | 43 / 43 | 43 / 43 |
| Opened a file outside the sales folder | 0 of 43 | n/a, returns passages |
| Top search result from outside the sales handbook | 24 of 83 searches (29%) | 44 of 54 searches (81%), all 43 that searched every pack |
| Tool output reaching the model per answer (median) | ~57,000 chars, ~14,000 tokens | ~9,000 chars, ~2,300 tokens |
| Traps where it said the sales handbook does not cover it | 13 / 14 | 14 / 14 |
| …and said so before answering | 8 / 14 | 13 / 14 |
| Wall time per answer (median) | 20 s | 17 s |
On correctness, this is a tie, and on scope Drive did better than we expected. Asked about the sales handbook, Claude searched Drive, found the folder, and opened only files from it on all 43 real questions. Unrelated files from the same Drive showed up in its search results for 23 of the 43 questions, but Claude never opened one.
Drive's search returns long snippets of each matching file, and its connector then reads a whole file. For each question, Claude searched, opened the pages that looked relevant, and read each one in full. Across the 43 real questions that came to a median of about 57,000 characters of tool output per answer: about 27,000 of search results and about 25,000 of file reads.
A Warmhand search returns the few sections that match, from whichever packs hold them. The median was about 9,000 characters. That figure counts everything the pack's tools returned in the chat, the formatting around each result and the occasional second search included. Of it, about 2,900 characters were passages from the sales handbook, about 3,600 were items from other packs, and about 1,800 were the list of packs that matched. It is higher than the 665 tokens per search we reported in post 2, for that reason, and because these searches ran across everything the account holds, not one pack.
Both figures come straight from the transcripts. Tokens are characters divided by four.
Claude's first search in every chat covered everything the account holds, without naming a pack. The account we tested in is a working one, with other packs in it besides the three handbooks. In all 43 of those first searches, the top-ranked result was a whole item from one of those unrelated packs. The same item came first for 14 different questions. The right sales passage was in every one of those results, usually second.
The 10 times Claude narrowed a search to the sales pack, the top result was a sales passage every time. Drive's top search result was outside the sales folder in 24 of its 83 searches.
The answers were still right, because Claude read past the first result. But a search that puts the right passage second is not good enough, and more than half of what we handed the model was material it did not need. On our test stack, which held only the three handbook packs, the top result came from the wrong pack 2% of the time. A real account is the harder case, and ranking across many packs is what we are fixing next.
Part of the cause is the same as in post 2: Claude never made the call that tells it which packs exist, so it rarely knew to narrow a search to the sales pack.
The traps are where the two sides behaved differently.
Asked "How long can a PagerDuty alert go unacknowledged before the on-call manager gets alerted?", both sides found the answer in the support handbook. Warmhand opened with the source: "The sales handbook doesn't cover this; it has nothing on PagerDuty escalation. The answer is in the support handbook." Drive opened with the answer, "10 minutes", and named the support document a few sentences later.
On 13 of the 14 traps Warmhand said the sales handbook did not cover the question before giving the answer, and on the 14th it said so at the end.
Drive said it on 13 of 14 as well, but before answering on only 8. Five times, the PagerDuty question among them, it gave the support handbook's answer first and mentioned the source afterwards. On a question about account lockouts it never said the answer was not from the sales handbook at all, only that it came from a file in the person's Drive.
For someone who asked about the sales handbook, that ordering matters. An answer that leads with a number from another team's rules reads as the sales rule.
Median wall time per answer was 20 seconds for Drive and 17 for Warmhand. Both include the model's thinking and page loads. In this account Drive's tools were already allowed to run; we approved each Warmhand tool call by hand, 103 clicks over 58 chats.
Who needs access to what. Drive answers from files the person can open, so everyone who should get the handbook needs Drive access to it, and the AI sees the rest of their Drive too. A pack carries the copy someone chose to share, and the person receiving it needs no access to the original files.
Where it works. The Drive connector is a Claude feature. The packs answered these questions in claude.ai, and the same packs connect to Claude Code, Cursor, ChatGPT or any other tool that speaks MCP.
What it costs per answer. Even with the noise above, the pack handed the model about a sixth of what Drive did, on every question. That leaves room for the actual work.
If your team's Drive is messier than ours, reply and tell us how. A real one is a better test than one we built.