Here’s a trick or a tip that I found out can help in most cases, mostly at the early stage of the problem but can also work otherwise as well but you’re most likely to get a higher success rate using this at the initial phase of an issue.
Sometimes you tend to notice an error and you try prompting the agent to fix it, the agent might fix something but maybe not what you We’re trying to fix
And you might find out continuous prompting to fix the issue might get the agent stuck in a loop
What I found helpful was
When you discover an error you wish to fix
Explain what’s wrong+happening to the assistant and assistant will give you details on what’s wrong(it’s going to suggest changes, don’t apply them, you are likely to get the agent confused when it to do with major code changes)
What you now have to do is copy the report and and paste it to the agent
Now the agent specifically knows where to look at and fix the error (this is a more targeted approach I think)
Hopefully this made more sense than it sounding alright in my head
Great call! It also seems like there’s been an update of some kind (or maybe my Agent just coded this) where when I get an error or the page doesn’t load, there is an explicit error log which I then feed back to Agent and it’s able to respond much more quickly. It’s now taking ~10-15 min of debugging what previously was taking hours of back-and-forth previously
Thanks for the insights. I am stuck for this past week implementing stripe payment. Well, I will try your new approach and see. This feed showed up at my notifications as someone is looking over my shoulder. What a coincidence. I love you all a Replit, keep it up🎉
Mine got stuck with a push command, giving an outdated command, I got the correct syntax and pointed that out, and after a couple tries it finally learned the right way
I’d like to note that sometimes I use an extra layer and it seems to work well. Example, I’ll go to X.ai, explain what I’m trying to do, have it make a prompt, then show it to Assistant to refine it further, then if I’m not sure I’ll go back to X to double check, then I’ll feed the prompt to Agent when I feel like it’s ready.
Would be good to be able to load context into a custom GPT to save time but I think Sam Altman is too evil so I don’t like to pay for ChatGPT lol. Then again I pay for Office so maybe I should stop being stubborn. Stupid inconsistent meat brain!!
