Power Platform Community Forum Thread Details

I’m building a Copilot Studio agent that creates Word (.docx) files from natural language. When the content is large, the generated file shows strange characters like â€“, Ã, Ã© instead of proper dashes and accented letters (Portuguese text).

I tried reopening with UTF‑8, using Notepad/Notepad++, and paste options, but nothing worked.

Question:
How can I ensure the Word files generated by Copilot Studio use the correct encoding (UTF‑8/Unicode) and avoid these corrupted characters? Any best practices or recommended approach?

Categories:

Bot administration

Building Copilot Studio chatbots in Microsoft Teams

General topics

Calling actions from Copilot Studio

hello,

i have come throught multiple document creation : image, excel, power point. To be honest, word will be my next try next week.

What i can share you throught all my test.

I recommand you to use the Prompt from tools (AKA Ai Builder prompt), enable code interpreter and then go to GPT 5 Chat (not reasoning, it will timeout)

I don't recommand from copilot studio directly, i recommand from AI Builder prompt and use it inside a flow, this way you will

- test in AI builder,have a very big prompt windows, and when the result is stable, the generated code will be lock when clicking on save button, this way you will 100% accurate result every time

- since it will generate python : use another AI tool to help you like copilot or last version of GPT 5.2 wich is good enough to help in this case : explain the case, give the desired file, give the result, the prompt, the generate code and then the AI will help you build a prompt for AI builder wich all request abouht the code generation to avoid this kink of problem

- then save and use an agent flow to get the file in base 64 and save it wher eyou wan t (often in one drive of the user, it s great to keep track)

i recommand you this content wich is with excel but quiest near : https://www.youtube.com/watch?v=JMlUcxqnU30

last suggestion : when i made pptx generation i found myself that the python was very sensitive of the quality of the hidden layout of the pptx. I think for word it will be nearly same, the police layout (title 1, subtitle 1 etc) will need to be perfect too.

if this help you to resolve the case, please mark the answer as verified, it's important for the community :)