dragonscave.space is one of the many independent Mastodon servers you can use to participate in the fediverse.
A fun, happy little Mastodon/Glitch instance.

Server stats:

238
active users

If you're a user and struggling to find the start of responses in the interface, you can instruct it to add a heading at the start of each one:

1. In the web interface, press the "User" button.
2. In the menu, choose "Customize ChatGPT".
3. In the dialog that comes up, the second multiline text field is labelled: "How would you like ChatGPT to respond?". In this field, you can add something like: "Whenever you respond, prepend a level 1 heading in the Markdown so I can easily find the start of the response with my screen reader.".
4. Near the end of the dialog, ensure that the toggle switch is in the on position, which will confusingly mean it has a label of: "Disable chat preferences". But this does mean it's enabled.
5. Press "Save".

In my experience, this works for something like 96 to 98 percent of responses, and is probably possible with other models too.

A few responses I'm not interested in:
1. I hate AI.
2. I hate OpenAI.
3. I hate ChatGPT.
4. OpenAI should be better at accessibility (true but unhelpful).
5. "This is a level 1 heading, should it be level N?"

simon.old
Public

@jscholes This is actually super clever. I wrote to them about an API issue last year and got "please clear your browser cookies and try again" as a response...maybe they should put their own AI model in charge of support. I think it'd do a better job. But in the meantime, this is a really good DIY solution and something that would never occur to me. It could probably also write a whatsitmonkey script to fix the various issues permanently, but I haven't tried this.

James Scholes
Public

@simon The nice thing is that in cases where the first line of the response is short, it just makes that the heading. In cases where the line is longer, and hence unsuitable to be a heading, it seems to use the same sort of approach to titles that is used for the chat history listing. No idea if I could shorten the instructions to save tokens while still being effective, but as this isn't via the API I don't care much.

Tim Ward ⭐🇪🇺🔶 #FBPE
Public

@simon @jscholes "Clear your cookies/cache and start again" is sometimes useful advice if the authors of the web application haven't worked out that they should version such things. But then ... if they haven't worked that out, how do they know to give that advice?

simon.old
Public

@TimWardCam @jscholes It's useful advice if you're actually using a web application. I wrote to them about an issue with the Whisper API. There was no browser or cookies to clear. I also got the response almost a month after I sent in a request, so maybe support just got slammed and they threw the most common response at every single email. F for effort.

Jeffrey D. Stark
Public

@simon @jscholes I wonder if there's a way to do this with copilot

Leo
Quiet public

@jscholes Oh interesting, I didn't know it could do that. I remember asking it once whether it can modify the html format as it responds, and and it told me it couldn't. This was before custom instructions, so this is great to know!

Serena 🏳️‍🌈
Public

@jscholes @simon OMG thank you, this is so handy. i’ve done this now.

Leo
Quiet public

@sapphireangel @jscholes One thing to keep in mind is that it'll follow those instructions even when you use voice mode. So instead of just responding to your input, it might also read the heading out loud. I told it not to do that during voice mode and it sort of works, at least 90 % of the time.