> You are in flow and the brain wants to move along but ai-inference becomes a bottleneck to the extend that we accept the garbage it gives as long the tests passes and loose the flow
The art lies in being in the flow at an altitude higher than the code. Once you discover that stream that exists at a higher altitude, the flow is even more exhilarating than that at the code level. It is almost liminal.
If you use open weight models, you can run several agents for the same budget. From what I've seen of Claude, there is no difference for most daily work. Claude makes many mistakes too. The real thing that sets open weights appart is that you can leverage multiple model families from the same token vendor. Using multiple families has outsized impact imo. Like... don't have Claude review the code it wrote.
I time slice between them, have them do a lot more reviews of each other, and spend more effort on thinking about design and checking their code.
I'd argue doing audits and having agents clean up their own slop, then creating instructions / guidelines from those, will reduce the problem long-term. They very much will mirror the surrounding code the code is consistent itself and the core points are called out in a little AGENTS.md
The question is, who is ordering you to work this way, and why can you not tell them to fcck off?
> You are in flow and the brain wants to move along but ai-inference becomes a bottleneck to the extend that we accept the garbage it gives as long the tests passes and loose the flow
The art lies in being in the flow at an altitude higher than the code. Once you discover that stream that exists at a higher altitude, the flow is even more exhilarating than that at the code level. It is almost liminal.
If you use open weight models, you can run several agents for the same budget. From what I've seen of Claude, there is no difference for most daily work. Claude makes many mistakes too. The real thing that sets open weights appart is that you can leverage multiple model families from the same token vendor. Using multiple families has outsized impact imo. Like... don't have Claude review the code it wrote.
I time slice between them, have them do a lot more reviews of each other, and spend more effort on thinking about design and checking their code.
I'd argue doing audits and having agents clean up their own slop, then creating instructions / guidelines from those, will reduce the problem long-term. They very much will mirror the surrounding code the code is consistent itself and the core points are called out in a little AGENTS.md