Make files and use skills in chat
Ask the assistant for a document, spreadsheet, presentation, chart, or image, and let it follow a skill from your library, without leaving the conversation.
5 min read
The assistant can make what you ask for in the conversation: a CSV or Markdown file it writes itself, a Word document, Excel workbook, PowerPoint presentation, PDF or chart it builds in a light sandbox, and images when your organization allows them. The files appear below the reply, ready to open and download. Skills from your skill library tell it how to make them well.
Ask for the file you need
Name the format and the content in your message, as you would to a colleague. For example: “Make an Excel file with our monthly revenue for 2025 and a total row that uses a formula.” Or: “Turn this into a nicely formatted PDF,” followed by the text.
The assistant uses the lightest way that does the job, and starts nothing it does not need:
| You ask for | The assistant | What it starts |
|---|---|---|
| An answer, a comparison table, a short code block | Answers in the reply | Nothing |
A Markdown, CSV, JSON, HTML or SVG file, a calendar entry (.ics) or a contact card (.vcf) | Writes the file itself | Nothing |
| A picture: an illustration, a photo-like scene, a logo | Generates an image | Your organization’s image model, when image generation is on |
| A Word, Excel, PowerPoint or PDF file, a chart, a calculation over data, a change to a file you attached | Runs code in a light sandbox | A light sandbox, in about a second |
| Work on a repository, in a browser session, or lasting hours | Suggests a project task | Nothing in the chat; the task’s agent works in its own sandbox |
A file that should be reviewed, kept with other work or made later belongs on a task: the project's standard agent uses the same tools there, and the files it makes become the task's deliverables.
While it works, the steps above the reply say what it is doing, such as Writing contacts.csv, Reading the xlsx skill or Running Python in a light sandbox. A failed step says what went wrong, and the assistant usually corrects its code and tries again.
Open and download the files
The files appear under the reply, in the order they were made. Images show as pictures; select one to open it full size and page through the others. Every other file shows as a card with its name, type and size. Select the card to preview the file; PDF, Word, spreadsheet, Markdown and text files open in place. The download button beside the card saves the file under its own name.
When the assistant revises a file in the same reply, for example after checking its own output, the card shows the final version only. Ask for a change in a follow-up message to get a new version in the next reply.
Files belong to the conversation: whoever can open the conversation can open its files.
Change a file you attach
Attach the file and say what should change: “Add a chart of the totals to the attached spreadsheet,” or “Convert the attached Word document to PDF.” The assistant copies the file into the light sandbox and returns the result as a new file. The file you attached stays as it was.
Use a skill
A skill is a tested procedure for a kind of work, such as a document format or a house style. The assistant sees the name and description of every skill in your skill library: those shared with the whole organization and with your teams. When one fits your request, it reads the skill before it starts and follows it; the steps then show, for example, Reading the docx skill. The built-in skills docx, xlsx, pptx and pdf cover the Office and PDF formats.
To be sure a skill is used, name it: “Use the brief-summary skill to summarize the attached brief.” A skill whose author marked it for explicit use only is used only when you ask for it by name. In a library of more than 40 skills, the assistant sees the first 40 in alphabetical order and uses any other one when you name it.
A skill guides the assistant; it does not change what the assistant is allowed to do. A skill’s scripts and templates run in the light sandbox, like the assistant’s own code.
Know the limits
- A light sandbox belongs to one conversation. It keeps its files between the assistant’s steps and your follow-up messages, and after about five minutes without use it is removed; the next request starts a fresh one. When your organization’s light sandboxes are all in use, an idle one gives way; when none is idle, the assistant tells you to try again in a minute.
- One run of code may take up to three minutes, and the tools of one step may run for at most five minutes together. At most 16 tools run in one step. Press Stop generating and a running program is cut off within seconds. A run attaches up to 10 files, each up to 20 MB; when a program writes many more, the reply names a few and counts the rest.
- A file the assistant writes itself holds up to 1 MB of text.
- One request creates up to 4 images, and one reply up to 16. Images are billed to your organization and count against you, like the reply itself.
An administrator decides what chat can make under Chat tools, whether it can create images under Image generation, and how many light sandboxes your organization runs at once under Sandboxes. A tool that is switched off is gone from every chat; the assistant then answers in the reply or suggests a task instead.