---
title: Giving it materials
description: Three ways to hand over files, what it accepts, the size limits, and the one difference between a chat and a handed-off task.
---

There are three ways to hand a file over, and they all end up in the same place:

- **Attach files** in the composer, and pick them.
- **Paste** into the composer. Pasted screenshots and clipboard images work too;
  they get named for you.
- **Drag and drop** onto the composer.

Nothing else is needed — no upload step to complete first, no folder to organise.

## What it accepts

- Plain text and Markdown, CSV, JSON, and common source files
  (`.ts`, `.tsx`, `.js`, `.jsx`, `.py`, `.html`, `.css`, `.xml`, `.yaml`, `.yml`)
- PDF
- Word (`.doc`, `.docx`), Excel (`.xls`, `.xlsx`), PowerPoint (`.ppt`, `.pptx`)
- Images — PNG, JPEG, GIF, WebP

Anything outside that is refused at the door with a plain reason rather than
accepted and quietly ignored: *"…isn't supported. Add a document, spreadsheet,
image, PDF, text or source file under 5 MB."*

## The limits

**Five files per run, 5 MB per file.** If you have more than five, either send
the five that matter or combine them first. If a file is over the limit, exporting
the relevant sheet or pages usually gets it under.

## Chats take more than handed-off tasks do

This is the one asymmetry worth knowing before you plan a job:

| | Attached material |
| --- | --- |
| A chat | Up to five files, any supported type, 5 MB each |
| A handed-off task | Up to five files, **plain text only**, 100 KB in total |

The reason is where the work happens: PDF, Office and image inputs are handled
in the live chat so they can be mounted in the secure workspace. So if
the job depends on a PDF, a spreadsheet or a screenshot, do it in a chat.
Hand off a task when the material is text — a list, a brief, a pasted export.

The check happens on both sides, so a task that exceeds this is refused when you
create it, not halfway through the run.

## Say what each file is

Files arrive without their story. *"The first one is last quarter's actuals, the
second is the forecast I'm trying to correct"* costs you eight words and removes
the most common failure — correct work performed on the wrong file. Where two
files disagree, say which one wins.

## Scanned documents

It can pull the text out of a PDF, convert most files into clean editable text,
and read and write Word documents, spreadsheets and slide decks. Separately from that, some
makers document their own model as taking a PDF file directly rather than text pulled out of
it first — [models whose maker says they can read PDFs](/docs/models-that/read-pdfs) lists
the ones you can pick that carry the claim, and is explicit that it is their claim rather
than something we measured.

What it cannot do is read a scan. A PDF that is only page images with no text
layer yields nothing, and in that case it tells you the document came back empty
instead of inventing contents that sound like your document. If that happens,
send the original file rather than a scan or photo of it, or paste the relevant
text in directly.

## Things you keep re-attaching

If the same material is really an instruction — a format you always want, a
checklist you always follow — teach it once instead of attaching it every time.
That lives under **Things you taught it**, on the Assistants screen, and is
reusable in any chat afterwards.