Summary and understanding of Canvas issues | OpenAI | Page 1

crystal shoal Oct 10, 2024, 10:33 PM

#

Many issues with canvas can be understood with the following:

Canvas supports multiple "documents" in a single context/session.

We create a document with a prompt. The response creates a document placeholder which includes a document name, type, and a reference to the content.
The name is based on the name that we provide. This is not a file name, it's a tag for reference.
If we upload a file the type is inferred but it's often wrong. (bug) We need to be explicit in our prompt about what kind of file is being used. This sets internal metadata like "code/tsx" which is later used for syntax highlighting.

Sometimes the document type is incorrect and we can't change it. (issue)

To populate the canvas we must request a document by (approximate) name and use phrases like "show the Foo document in the canvas" or "open Foo". Combining this request with others, like "open Foo and change ..." often results in bad changes. (bug)

A document can be updated even when off-canvas. This is a feature but may confuse the model.

The current document in canvas is often forgotten, causing the loading of a "doc2" loading into canvas with the header "doc1", and corrupting doc1. (bug) The model may insist doc2 is loaded when we tell it doc1 is loaded. (bug)

We can't create an empty named document. New documents must have starting content or it breaks. (bug) I address this with this prompt: "Create a new document named App of type TSX which has a single comment // as content."

An attempt to create multiple documents with a single prompt usually results in a "Oops" error. (bug)

In summary , the general issue is that the interface gets confused with creating, referring to, and updating document objects. Multiple documents in context, switched in/out with canvas, get confused. This isn't only caused by context size but that may be a factor.

I hope that helps OpenAI to address related issues and that it helps other users to understand the issues.

elfin krakenBOT Oct 10, 2024, 10:33 PM

#

@crystal shoal

Please follow the posting guidelines.

For better assistance from the community or an OpenAI team member, please follow the posting guidelines by clicking the button below.

crystal shoal Oct 17, 2024, 3:57 AM

#

The canvas interface seems to be better today but still has some of the same faults. I dunno if there have been changes or if I'm just not tripping on the issues. Some word in a change log somewhere would be helpful.

crystal shoal Oct 31, 2025, 5:13 PM

#

I opened this thread over a year ago and the same issues persist. I haven't seen any improvements in Canvas in the last year. Today I tried it again to see how it might help with revising a README document for a FOSS project. The attached screenshot is a simple example of how even GPT-5 can't properly manage a Canvas document. To be clear, knowing that Canvas doesn't work with multiple documents I pasted a README file into a single document and asked the assistant to break it up into pages using a double literal HR tag, rather than triple dashes, which only clashed with existing use of triple dashes. It wiped out all content except headers, eliminating any work that it had already done. This is not unexpected. It's typical of how messed up the Canvas feature is. I'm just publishing this here as evidence of an ongoing issue with this potentially great but as-yet incomplete feature.

#Summary and understanding of Canvas issues