Retry but with instruction to AI including the failed message

[Thoron77]

I was struggling with taming MiniMax to follow instructions and got an idea ;)

How about adding a retry but with system instruction with optionally the failed message and a prompt how AI should correct it, like check with given instructions, or make sure to include this and that...

Similar functionality would be Undo couple times and add author's note with instruction but it is much less precise because there is no way to show AI where it is failing.

It could work similar to Add image with popup for prompt, where user can add out of context system/author's note prompt with instructions and option to include last message.

[LostRuins]

I don't really get it. Sounds like you're just looking for thinking models?

[LostRuins]

But how would the AI know it made a mistake?

[Thoron77]

This is the idea: When you do Retry, you add prompt and include last reply plus some background system instruction (like with Author's note for example). I'll think of how to illustrate it better and get back to you.

[Thoron77]

I'm so sorry it took so long but got busy, beside it is more difficult to illustrate than I thought. I'm realizing it may not be possible.
But here is more about the idea:

Lets say I have instructed AI to illustrate every reply with a , sometimes it just 'forgets' about it.
Now I have a generated reply which have no illustration added.

Also everything is configured correctly, Kobold is all capable and perfectly working generating pictures.

Now I hit the Retry or 'Retry with prompt' to differentiate those and I get the prompt:

Where I can add the prompt to instruct AI to take into account that instruction when generating reply which will overwrite the one already there. Optionally adding that last message.

How I picture it is like an Author's note in Context but just for this particular action.

The system prompt could be something along this lines:
Generate next reply but take into account this instruction: [value of text from regen prompt].
If ticked the include part:
Base your reply on text example corrected with the instruction. Text example: [Last message snipped]

Now that I've spent some time analyzing it all and trying to come up with this explanation I'm pretty sure its not a good idea as it will make the tracking messages tedious and overcomplicated.

Analogous to my idea would be to copy the last reply, Undo, add Author's note with immediate effect. In note add the prompt to correct the included text and paste the message. Then hit the generate reply button.

P.S.

Just for laughs I've clicked Retry and AI generated it as intended without any additional instructions just good old Retry ;)

[LostRuins]

I think it would be easier to just hit retry a few times... not sure its possible to correctly determine if the AI followed the prompt instructions or not

[Thoron77]

I agree, tried a few times with going back and adding immediate Author's note and got mixed results.