Prepare documents for Custom GPTs, NotebookLM, Claude Projects and Copilot

Custom GPTs, NotebookLM, Claude Projects and Copilot agents all answer from files you upload, and they work from the text in those files. This guide needs no code. It covers why a cleaned plain text file makes a better knowledge file, how to make one with Agnos or by hand, and what each tool’s own help pages say about their limits.

The short answer

For most documents, upload plain text that holds only the content: no page headers or footers, no page numbers, no website menus or cookie banners, and no second copies. All four tools accept .txt files.

With Agnos, the Upload pack for an AI tool export makes these files for custom GPTs, Claude Projects, Gemini Gems and NotebookLM: your cleaned documents as Markdown (.md, plain text with # headings), combined to fit the tool’s file limits. For Copilot, or any tool that wants .txt, use the Cleaned documents export: one .txt file per document.

Keep the original file when its meaning is in the pictures, such as charts or diagrams: a text file has none. Claude reads the images in a PDF of up to 100 pages, and ChatGPT Enterprise can search the visuals in PDFs, so a PDF can be the better choice there.

Why cleaned text files work better

• The tools read the text. OpenAI says that, outside ChatGPT Enterprise, ChatGPT takes the digital text out of a document and discards its images. Claude reads only the text of project files, apart from multimodal PDFs. Whatever is in that text, useful or not, is what they search.
• Clutter competes with answers. A PDF’s running header, a page number on every page, a web page’s menu and cookie banner: all of it becomes knowledge the assistant can draw on. A question can bring back the menu instead of the answer.
• The vendors ask for it. OpenAI’s help for GPT builders recommends clear, “text-forward” files, because complex layouts make content harder for a GPT to use. Microsoft advises keeping the files you give Copilot concise.
• You can check it. A text file shows exactly what the assistant will read. Open it before you upload it.

A real example: a saved web page

The sample project that comes with Agnos includes support-article.html, a help-center page saved from a browser. Around one short article, it holds this:

Before: the text around the article in the saved page

Home
Products
Support
Your account
Contact us
This website uses cookies to improve your browsing experience. Accept all
cookies, or manage preferences.
Please enable JavaScript to view this page.
  … the article …
Related articles
Moving your licence to a new PC
Which file formats are supported
Was this article helpful? Rate it out of five.
Terms | Privacy | Cookie policy

Upload the page as it is, and all of that becomes knowledge. Cleaned with the RAG / Retrieval Corpus purpose, Agnos’s Cleaned documents export holds support-article.txt, which begins:

After: support-article.txt from Agnos

# Why an export stops at the first file

An export that writes one file and then stops is nearly always a queue that never drained. The counter is still above zero, so the writer waits for work that has already failed.

## Before you start

Note the job number from the review screen. You will need it twice, and it is the only thing that ties a run to the files it read.

## Steps

- Open the job and read the queue counter at the top of the panel.
- If the counter is above zero, press Stop and wait for it to reach zero.
…

The menu, the cookie banner, the notice, the sidebar and the footer links are gone. The headings, the four steps and the code block at the end of the article, including a < b, are kept. The same cleaning takes running headers, footers and page numbers out of PDFs; see the headers and footers guide.

Make the files with Agnos

1. Start a project. In Projects, press New project and type a name. Check that the purpose is RAG / Retrieval Corpus, which suits documents an AI tool will search and answer from; if it is not, press Change and choose it. Press Create project.

2. Add your documents. Drop the files, or a whole folder, on the Import step. Agnos reads text-based PDFs, Word, PowerPoint and Excel files, web pages, CSV, JSON and plain text.

3. Look at what was removed. Open Review, choose a file and press Removed. Every removed piece names the rule that removed it.

4. Export an upload pack. Open Export and choose Upload pack for an AI tool. Under The AI tool you will upload to, choose your tool, or leave Any of them. Press Build and download.

5. Unzip and upload. Read READ ME FIRST.txt: it lists what is in each file and what Agnos took out of each document. Upload the .md files to your tool. Leave READ ME FIRST.txt and manifest.json out: they are for you rather than for the assistant.

For Copilot, or to get .txt files, choose Cleaned documents in step 4 instead, and upload the .txt files. What each file holds is in Export formats.

The Cleaned documents zip for the sample project, plus the handbook PDF from the other guides, held one text file per document:

northwind-documents_20260923_190359.zip
  faq.txt
  meeting-notes.txt
  price-list.txt
  staff-handbook.txt
  support-article.txt
  northwind-staff-handbook.txt
  manifest.json

Worth knowing

• Each file is named after its original document, or a combined file after the first document in it, so a citation points to a name you recognize. In the upload pack, each document also starts with its file name as a heading, so a combined file still says where each passage came from.
• The upload pack and Cleaned documents keep every copy of text that appears in more than one document, because two documents that share a paragraph are still two documents. To keep only the first copy, set Text found in more than one file to Keep one copy on Export.
• Upload one version of each document. An old and a new version side by side give the assistant two answers to choose from.
• A spreadsheet becomes rows of text, one sheet after another. If the assistant has to calculate with the numbers, a tool that analyzes spreadsheets may do better with the original file.
• If a tool limits how many files you can add, the upload pack combines neighboring documents to fit, keeping their order, and splits a very long document at paragraph breaks. READ ME FIRST.txt says which documents are in which file. If they need more files than the tool takes, it says to upload them in batches.
• Uploading sends the files to that company. Agnos cleans them on your computer, but it does not decide what is safe to share, so read them before you upload them.

Each tool’s limits

These come from each company’s own help pages, as of September 2026. They change, so check the linked page before you plan around a number.

The upload pack works to the limits the tools published in November 2026, listed under The AI tool you will upload to. If an upload is refused, check the tool’s help page.

Custom GPTs (ChatGPT)

• Up to 20 files per GPT, each up to 512 MB, according to Creating and editing GPTs. OpenAI’s File Uploads FAQ gives 10 files per GPT, so check the builder.
• Text and document files are capped at 2 million tokens each (File Uploads FAQ).
• OpenAI said on 11 September 2026 that it plans to retire custom GPTs in favor of plugins, and a GPT’s knowledge files carry over as reference files (retirement FAQ). Personal accounts (Free, Go, Plus and Pro) can no longer create new GPTs; Business, Enterprise and Edu workspaces still can for now (Creating and editing GPTs). Clean files matter either way.
• ChatGPT Projects take files too: up to 5 per project on Free, 25 on Go and Plus, and 40 on Edu, Pro, Business and Enterprise (File Uploads FAQ). On Free that is fewer than the 10 files the upload pack allows for ChatGPT, so check the number of files in READ ME FIRST.txt before you upload.

NotebookLM (now Gemini Notebook)

• Google renamed NotebookLM to Gemini Notebook in July 2026 (Google’s announcement).
• It accepts .txt, .md, .docx, .pdf, .csv and .pptx files, among other sources.
• Each source can hold up to 500,000 words, or 200 MB for an uploaded file, and free accounts can add up to 50 sources (Add or discover new sources for your notebook).

Claude Projects

• Project files can be up to 30 MB each. There is no fixed number of files, but together they must fit in Claude’s context window, and Claude reads only their text, apart from multimodal PDFs (Upload files to Claude).
• On paid plans, when project knowledge nears that limit, Claude switches to retrieval (RAG) and can hold up to 10 times as much (What are projects?).
• Accepted documents include TXT, PDF, DOCX, CSV, HTML and JSON.

Copilot (Microsoft)

• Agent Builder in Microsoft 365 Copilot takes up to 20 files uploaded from your device: .txt, .docx, .pdf and .pptx up to 512 MB each, and .xlsx up to 30 MB. Microsoft advises keeping the files concise (Add knowledge sources to your declarative agent).
• Copilot Studio takes up to 500 files per agent, each up to 512 MB, and accepts .txt and .md among many types (Upload files as a knowledge source).

Markdown (.md) is not on every list, but plain text is. The upload pack’s files are .md, made for ChatGPT, Claude, Gemini and NotebookLM. Cleaned documents gives .txt files, for Copilot and any other tool.

Without Agnos

For a few documents, you can make clean text files by hand:

• Word: choose File, Save As, and the Plain Text (.txt) type. Open the result and check it.
• PDF: copy the text into Notepad, delete the header, footer and page-number lines, and save it as .txt.
• Web page: open it in your browser’s reading view, if it has one, copy the article, and save it as .txt.

That is fine for a handful of files. For dozens it is slow, it is easy to miss a line or delete a real one, and nothing records what you took out.