Rubric
Contents — domains, guide and mocks

Managing knowledge and connectors

CCAO-F 5.216 min read · checked 21 September 2026

Task statementManage uploaded knowledge and connectors (e.g., Google Drive, Gmail)

Uploaded copy versus connected source

Uploaded file

  • A snapshot taken at the moment of upload
  • Works with no integration and no admin involvement
  • Goes out of date silently — nothing announces it
  • You control exactly what Claude has: this file, nothing else

Connector

  • Reads the source as it is now
  • Mirrors your existing permissions in that system
  • Needs organisational enablement on Team and Enterprise
  • Can take actions, not only read — which admins can restrict
Neither is better in general. The question is whether the information changes, and whether a permission boundary matters.

Uploaded knowledge: know the limits before you promise anything

Uploads land in two different places with two different limits. A file uploaded into a single chat can be up to 500MB, with up to 20 files per chat. A file added to a project’s knowledge is capped at 30MB, and while the number of files is not limited, the total content has to fit within Claude’s context window — beyond which, on the paid plans, retrieval takes over. Supported document formats include PDF, DOCX, CSV, TXT, HTML, ODT, RTF, EPUB and JSON, with XLSX requiring code execution and file creation to be enabled. Images can be JPEG, PNG, GIF or WebP, up to 8000 by 8000 pixels.

WherePer-file sizeHow manyNote
A single chatUp to 500MBUp to 20 filesGood for one-off analysis of something large
Project knowledgeUp to 30MBNo stated capTotal content must fit the context window
PDF, up to 100 pages——Text and visual elements such as charts are analysed
PDF, 101 to 1,000 pages——Text only; visual elements are not analysed
DOCX and similar——Text extracted; embedded images are not interpreted

Those last two rows cause more silent failures than any size limit. A 400-page PDF whose crucial numbers live in a chart is, for practical purposes, missing those numbers. A Word policy whose approval matrix is a pasted screenshot has no approval matrix as far as Claude is concerned. Neither produces an error message; you get a fluent answer built from whatever text was available. The habit that catches it is asking Claude to quote the passage behind an answer — a quote that cannot be produced tells you the content was never read.

Where should this knowledge live?

How does this information behave?
  • Stable, needed often
    Project knowledgecurrent version, clear filename
  • Changes weekly or daily
    Connectorread the live source instead
  • One task, then done
    Chat uploadno configuration to maintain
  • Restricted or out of policy
    Neithercheck with the data owner first
The two questions that decide it are how often the content changes and how many chats need it.

Connectors: what they are and who must switch them on

Connectors let Claude reach your apps and services — retrieving data and, where permitted, taking actions within them. Anthropic’s documentation divides them into directory connectors, which are pre-built and verified integrations for well-known services, and custom connectors, which point at a remote MCP server at a publicly reachable address. You manage them under Customize and then Connectors, or reach them from the chat’s menu, and you can toggle an individual connector on or off for a particular conversation.

Two rules govern what a connector can see. It inherits your permissions: Claude cannot access anything you could not already open in that system. And on Team and Enterprise plans, an owner has to enable connectors for the organisation before individuals can authenticate at all; admins can also restrict what actions are allowed — read-only rather than write, for instance — and can prevent connections to services outside the organisation’s verified domains.

A connector request, end to end

You
Claude
The connected service
Step 1: You to Claude: “Find last week’s pricing memo”
Step 2: Claude to The connected service: Search, as you, with your access
Step 3: The connected service to Claude: Only files you can open
Step 4: Claude to You: Answer, citing the file
The permission check happens in the source system, not in Claude. That is why a connector cannot leak a document you were never allowed to open.

The Google Workspace family is the set the objective names, and it is three connectors, not two. Gmail can search and read your mail, draft and send messages, manage labels and read message metadata. Google Calendar can see events across your calendars, create and update them, and find mutual availability. Google Drive can search and retrieve documents in several formats, share and move files, upload files, and save Claude-generated files straight back to your Drive.

ConnectorUseful forA limit worth knowing
GmailFinding a thread, summarising a customer’s history, drafting repliesAttachment contents are not accessible — only message metadata
Google CalendarAvailability, preparing for a week, chasing responsesWorks within your own calendar permissions
Google DrivePulling the current version of a document into a chat or a ProjectEmbedded images in documents are not processed

Two further constraints shape design. Connectors are subject to the underlying provider’s rate limits, so a request that would sweep thousands of items is not a good pattern. And there are two distinct ways to use Drive material: adding a file directly to a conversation — by pasting its URL or picking a recent document — or adding it to a Project’s files, which Anthropic notes is available when adding to files in private projects. Data retrieved through a connector is retained with the chat it appears in, so deleting the chat deletes that retrieved data, and it is encrypted in transit and at rest.

Curating what is in there

A knowledge base is a garden, not a warehouse. Documents are added by people trying to be helpful, superseded by newer versions nobody removes, and eventually contradict each other. Because retrieval finds the passage that best matches a question, an obsolete document with unusually clear wording can outrank the current one. Curation is therefore not tidiness — it is accuracy work.

A knowledge audit

  • Fails: One current version of each documenttwo versions of the standard contract are present
  • Check: Filenames carry subject and datethree files named “final”
  • Passes: Fast-changing data comes from a connector, not a copyrate card via Drive
  • Check: Formats are ones Claude can actually readone key table is an image in a DOCX
  • Passes: Nothing is in here that policy forbidschecked with the data owner
  • Passes: Connectors are enabled and still authorised
  • Missing: A named person owns this knowledge basenobody has been asked to own it
This takes twenty minutes a quarter and prevents the class of error that is hardest to spot: a correct answer from an outdated source.

Traps the wrong answers are built from

Tempting but wrongDo this instead
Treating an uploaded file as if it were the live sourcePut fast-changing data behind a connector, and date the filenames of anything you keep as a copy.
Assuming a connector gives Claude everything in the systemConnectors mirror your own permissions and have documented gaps, such as Gmail attachment contents.
Leaving superseded documents in project knowledgeKeep one current version; retrieval cannot tell which of two versions governs.
Assuming a document was fully readCheck the format rules and ask for a quote; long PDFs are text only and embedded images are not interpreted.
Connecting a service without checking policy or admin settingsConfirm the data owner is content and that an owner has enabled the connector for the organisation.

You should now be able to

  • Choose between a chat upload, project knowledge and a connector for a given piece of information.
  • State the file size, count and format limits that apply in chats and in projects.
  • Explain what Google Drive, Gmail and Calendar connectors can and cannot do, including the permission model.
  • Identify the silent failures caused by long PDFs and embedded images, and test for them.
  • Audit and curate a knowledge base so retrieval cannot surface superseded material.
  • Check organisational enablement and data-owner approval before connecting a service.

Practice questions

Original questions written for this lesson, in the exam’s style. Answer first, then open the reasoning — every option is explained, including why the wrong ones are tempting.

  1. Question 1

    A sales team keeps its price list in project knowledge, re-uploading it whenever someone remembers. A rep quotes a price that changed three weeks ago; the answer cited the uploaded file and sounded entirely confident.

    What is the best correction?

    1. AAdd an instruction telling Claude to warn when information might be outdated.
    2. BRe-upload the price list every Monday as a standing task.
    3. CRead the price list through a connector so Claude sees the current version.
    4. DMove the price list from project knowledge into each chat as an upload.
    Show answer and reasoning
    1. AIncorrect. Claude has no way to know the file is out of date; the instruction produces a warning that is either constant or absent.
    2. BIncorrect. It shortens the window but keeps a manual step that will be missed, and nothing shows when it has been.
    3. CCorrect. Fast-changing data belongs at its live source; a connector removes the snapshot problem rather than shortening it.
    4. DIncorrect. It is the same snapshot, now repeated per chat, with more chances for someone to use an old copy.

  2. Question 2

    An analyst adds a 350-page technical PDF to a Project. Much of the important information is in charts. Claude answers questions about the document fluently, but two figures it gives cannot be found anywhere in the text.

    What is the most likely explanation?

    1. AThe file exceeded the 30MB project limit and was only partly uploaded.
    2. BPDFs over 100 pages are processed as text only, so the charts were not analysed.
    3. CProject knowledge cannot hold PDFs, only text files.
    4. DThe analyst needs to enable code execution to read PDFs.
    Show answer and reasoning
    1. AIncorrect. A file over the limit would be rejected rather than partially ingested, and size is not what governs chart reading.
    2. BCorrect. Above 100 pages, visual elements are not analysed, so chart-only figures are absent and may be filled in plausibly.
    3. CIncorrect. PDF is a supported format for project knowledge; the constraint concerns how long PDFs are processed.
    4. DIncorrect. Code execution is required for XLSX spreadsheets, not for reading PDFs.

  3. Question 3

    A team on an Enterprise plan wants to use the Google Drive connector to pull current policy documents into their Project. One member connects successfully; three others see no Drive option at all.

    Which TWO explanations should be checked first? (Select 2.)

    1. AAn owner has enabled connectors for some users or groups but not all.
    2. BThe three colleagues have not completed the authentication step for the connector.
    3. CThe Drive folder is too large for the connector to index.
    4. DThe documents are in a format Claude cannot read.
    5. EProject knowledge must be emptied before a connector can be used.
    6. FClaude cannot use Drive and Gmail at the same time.
    Show answer and reasoning
    1. ACorrect. On Team and Enterprise plans, organisational enablement governs who can authenticate a connector at all.
    2. BCorrect. Each user authenticates individually, and until they do the connector is not usable for them.
    3. CIncorrect. Connectors search the source rather than indexing a folder in advance, so folder size is not the blocker described.
    4. DIncorrect. A format problem would produce poor answers, not a missing connector option in the interface.
    5. EIncorrect. Uploaded knowledge and connectors coexist; one does not exclude the other.
    6. FIncorrect. The Google Workspace connectors are separate and can be enabled together.

  4. Question 4

    An HR manager asks whether connecting her Gmail will let Claude analyse the CVs candidates have sent as attachments over the past year.

    What is the accurate answer?

    1. AYes, the connector reads mail messages and their attachments together.
    2. BYes, provided the whole mailbox is under 500MB in total size.
    3. CNo — Gmail exposes metadata, not attachment contents; the CVs need another route.
    4. DNo, because connectors may never be used with personal data of any kind.
    Show answer and reasoning
    1. AIncorrect. Anthropic documents that attachment contents are not accessible through the Gmail connector.
    2. BIncorrect. The 500MB figure is a per-file chat upload limit and has nothing to do with connector access to attachments.
    3. CCorrect. It states the documented limit and points toward uploading the files or reading them from Drive instead.
    4. DIncorrect. Policy may restrict this use, but the product limit here is about attachments rather than a blanket ban.

Practise every question in Domain 5 →

Sources

Drafted with AI assistance and checked against the sources above; expert review is in progress. Spotted an error? Tell us and it gets fixed, dated and listed on how this is written.