Replies: 7 comments
-
You can use just a simple text file with an unordered list in order you want the LLM to work. I actuall did something similar and also included these lines after gpt was happy to fully uninstall and reinstall npm after it messed up a file :D :
|
Beta Was this translation helpful? Give feedback.
-
How do you provide this file to the agent? Guide him to read it first? |
Beta Was this translation helpful? Give feedback.
-
Is it possible to keep certain instructions in a hard context which will not be overriden? |
Beta Was this translation helpful? Give feedback.
-
Yes, just tell it open the file xxx in folder yyy. If you specify the correct folder, you can save some roundtrips. There is no hard context except the system prompt by OpenDevin, which is not (yet) editable. This is a real message from gpt-4o from just today, which made me 🤦 |
Beta Was this translation helpful? Give feedback.
-
Yeah, that's exactly what I'm doing right now, same results 😅 Very frustrating. Promoted him: pwd, use readmes, respect .gitignore when searching and he lets me down every time. I guess it will not change until the editable system prompt you mentioned. This can take it 10 levels up. |
Beta Was this translation helpful? Give feedback.
-
Editing the system prompt is just moving your requirements. The LLM itself is the issue. |
Beta Was this translation helpful? Give feedback.
-
@4tal I see you're asking lots of great questions on github issues! Joining the Slack channel and asking it there is also a good way to get these kinds of questions answered. The Slack channel is in the README of the repo. |
Beta Was this translation helpful? Give feedback.
-
Describe your question
I'm facing challenges when trying to provide consecutive instructions where the second instruction depends on the completion of the first. For example:
Read readme.md in path A.
Then, read readme.md of project B, which is specified in readme.md from instruction 1.
I'm unsure of the best approach to achieve this. Should I:
Give the second task first with a note indicating that the path can be found by looking at readme.md in path A?
Continue providing instructions sequentially as described?
I'm looking for guidance on the best practice for this use case. Any suggestions would be greatly appreciated.
Beta Was this translation helpful? Give feedback.
All reactions