Documents
Numbers say what happened. Documents say why: what a company does, what the board decided, which customer is at risk. Gatlas keeps both side by side, so an AI assistant can answer with the context as well as the figures.
Documents come in two layers
- Library holds the raw files synced from your Google Drive, as they are. You find it under Documents in the sidebar.
- Documents holds short, curated documents written from those files (and from your views): a company overview, a decisions log, the definitions behind your KPIs. These are what AI assistants read first.
Prompts, under Documents in the sidebar, is how the first layer becomes the second: a prompt reads a set of files and writes a document, by hand or on a schedule.
Sync files from Google Drive
Connect Google Drive to a company
On Sources, choose Add source, pick the company and Google Drive, and sign in to Google. Gatlas asks for read-only access and never changes or deletes a file in your Drive. You can also send a setup link to the person who owns the folders; see Connecting sources.
Pick the folders
Tick the folders to sync; the arrows open a folder to pick subfolders. If a folder is not listed, for example one in a shared drive, paste its link. Then save with the Sync button, which counts the folders you picked.
Check what arrived in the Library
Every file in a selected folder, subfolders included, lands in Library. Google Docs, Sheets and Slides are converted to text. PDF, Word, PowerPoint and Excel files are stored as they are and read when a prompt needs them. Files over 50 MB are skipped; hover a skipped file to see why. Files with no text in them, such as images, sync but count as Unreadable.
Note
Distill files into documents with a prompt
Admins and above create prompts. On Prompts, choose New prompt and fill in:
- Name and Prompt: what to write and how to lay it out, in plain language.
- Entity: the company whose files it reads. Narrow it with Connection or Folder prefixes when only some folders matter.
- Schedule: manual, @daily or @weekly.
- Max input tokens: a size limit on what the prompt reads. A run over it fails rather than silently reading less. Leave it empty for the default.
Open the prompt and choose Run preview to see the result on a small sample without saving anything. When it reads well, Run live writes the document, and every later run adds a new version. A prompt that fails three times in a row pauses itself; fix it and choose Resume.
Claude can also create prompts for you, including ones that read your published views or other documents alongside the files:
Create a weekly prompt for Northwind that reads the board meeting notes folder and keeps a decisions log: each decision, its date, and who owns the follow-up.
Publish a document so assistants can read it
A document is either Published or a Draft. AI assistants only see published ones. The first live run of a prompt creates its document as a draft, so you can read it before anyone relies on it. Publishing it is a request to Claude today; the dashboard has no switch for it yet:
Let Claude see the prompt
Edit the prompt and turn on Published to MCP, which is off for prompts created in the dashboard. MCP is how AI assistants such as Claude connect to Gatlas.
Ask Claude to publish its document
Over a read/write MCP connection, ask:
Publish the output document of the Northwind decisions log prompt.
A document is tagged to the company its prompt reads. On Documents, the Untagged count shows any that no scoped MCP connection can see.
Start with the documents that answer the most
On Documents, Ideas for getting started suggests the documents worth having early for each company, and shows which it already has. A few that pay off quickly:
- A company description: what it sells, to whom, and how.
- KPI definitions, so "active customer" means one thing.
- Accounting conventions, such as the fiscal year and currency.
- Data caveats and known gaps, so nobody trusts a number that should not be trusted.
- A decisions log.
Ask about documents in Claude
Once published, documents are part of what Claude can read. Ask for them directly, or ask a question that needs both:
What does the Northwind company description say about its largest customers, and how does that compare with its revenue by customer this year?
Claude reads the curated documents first. When none covers the question, it can search the raw files in the Library too; treat a search result as a lead to check, not a count. Documents follow the same access as views: an MCP connection only sees the companies it was given. See Getting the most out of Gatlas for how documents and views work together.