@visakanv@nosilverv This is almost right! Actually you can trace this type of incentive structure all the way back to nineteen ninety eight when the undertaker threw mankind off hell in a cell and plummeted sixteen feet through an announcers table.
The most important meta-tip that can help you figure everything out that everyone is overwhelmingly try to explain:
Tell Codex your overall goal, ask it how you can set it up for success, and encourage it to ask you questions. If you do that, it will tell you the tools it needs to be able to properly edit spreadsheets, and figure out with you any ambiguities.
You can dervive what to do just by thinking about how to best collaborate with another Thing
@leerob 5: Have your agent walk you through the changes it makes! Code reviews aren't the result of trying to find the best way to help a human *understand* what is happening and why. https://t.co/arSNxdrJKk
I build Walkthrough to do this, but you can also just ask your agent.
@trq212 Great prompt, and I've also been thinking about how to equip the agent to be able to more effectively communicate with the user.
Akin in spirit to your HTML use, I've created a Walkthrough skill that works in tandem with your prompt!
https://t.co/arSNxdrJKk
@pvncher@irl_danB I think the core improvement for workflows is that it in-harness empowers a frontier model to deeply plan for itself it’s own execution architecture. Eg I’ll need some agents to code, to review, etc.
All the other details, you can take it or leave it.
Strongly agree! I think I’ve invested a ton of time into this and I am confident in my skills relative to the field but also much more confident that I’m not near optimality.
I think some of this is individual skill, and also I think that the ecosystem of harnesses, skills, and other not yet named constructs is nowhere near optimality yet.
@WCNegentropy@vgrichina@DavidKPiano I agree it can go too far, but I think there isn’t any indication that happened here.
The main agent spending time getting overall context, and then sending a single subagent to do some work, is a good pattern when the eg tool calls of that work won’t help it going forward.