Model Generated Markdown Gets Rendered in UI (or is otherwise malformed)

iamjameswalters

I asked Claude Haiku to format a list of Markdown links with a prompt like this:

Convert these to a markdown list like the following:

- [MB-234 - Report Graph does strange things at times](https://linear.app/.../issue/MB-234/)

It looks like the markdown it returned got directly rendered by the Assistant UI:

I was able to work around this by selecting the Code Assistant, which rendered the output inside a codeblock.

Despite the fact that this sort of thing is typically a "coding" usecase, Markdown is broad enough that you might use it in a number of contexts, crafting forum posts etc., that I think Assistant should render a nice markdown codeblock (or at least escape it so that the UI doesn't render it) no matter what model you're using. (would help me use the smaller models for a task like this and save you money anyway!)

Rehan

It is hard to tell here if it is the model or the markdown processing. If you still have the thread, could you click the copy button on the assistant response and paste it here? That will show us if it is the model or our processing. Thanks 🙂

iamjameswalters

I don't have those threads anymore, but I made a few new ones with this prompt:

For everything in this list: google, bing, yahoo, ask

Create a list of markdown links like this:

- [Google](http://google.com/)

I asked Claude Haiku, ChapGPT 4.1 Mini, Gemini 2.5 Flash, and the Code Assistant, and this time, none of them initially provided me with a codeblock. But asking each of them to do so yielded a codeblock. So it looks like the markdown they return does get rendered by UI, but asking them to spit it out as a snippet makes them escape the output.

Interestingly, I decided to run this same prompt at claude.ai and chatgpt.com, and the same thing also happens at claude.ai. Only ChatGPT (that was 4o) managed to one-shot it.

I'm not sure how important a problem this is, that first day perhaps it just didn't occur to me to ask the model to follow up with a codeblock snippet. Since all the models seem to respond to that, this isn't such a big deal.

Rehan

Ah ok, just want to make sure I understand correctly; you want it to present it in a codeblock when asked directly for markdown? We use markdown for the general assistant output, so it sort of makes sense that its non-code output gets rendered. I suppose it is tough for the model to tell whether it should output in a codeblock or not - another user might expect the markdown to be rendered when it is asked to output markdown. So I think the best way to go about this would just be to specify that you want the markdown in a a codeblock.

iamjameswalters

Yeah, that's the idea. I didn't know that if I asked the other models to put it in a codeblock they would, so I don't think this is a huge deal.