@pvncher@modkin_mp Well I didn't measure that spending directly, that's fair, but I was just able to do more than before on single distributed task, basically some task was done 50+ times, if it helped here โ I don't see why can't it help reduce spending on a regular task
@pvncher@modkin_mp Most tasks can be decomposed for smaller models to handle, at least during planning. With single model I was out of quota in 2 days, once optimised I was able to fit in ๐ So the single model is expensive. These post are bs though, agree, I measured that: https://t.co/vzzzPPAXwN
Since Haiku 5.5 is out everyone started to care which model to run ๐ Recently I was doing a tricky migration across 50-ish microservices and using a single model was draining my weekly quotas in 2 days. Plain routing prompts everyone shares don't work. Let me show you what works
Since Haiku 5.5 is out everyone started to care which model to run ๐ Recently I was doing a tricky migration across 50-ish microservices and using a single model was draining my weekly quotas in 2 days. Plain routing prompts everyone shares don't work. Let me show you what works
@dr_cintas Plain prompts don't work. In my experience around 27% of agent/workflow launches were still lacking model setting and the choice they made was also not great. Here's how to fix this: https://t.co/vzzzPPAXwN
Since Haiku 5.5 is out everyone started to care which model to run ๐ Recently I was doing a tricky migration across 50-ish microservices and using a single model was draining my weekly quotas in 2 days. Plain routing prompts everyone shares don't work. Let me show you what works
There's also a dev kit embedded so you can collect your own before/after stats (in dev folder), just ask Claude to see the plugin source and run these scripts
Since Haiku 5.5 is out everyone started to care which model to run ๐ Recently I was doing a tricky migration across 50-ish microservices and using a single model was draining my weekly quotas in 2 days. Plain routing prompts everyone shares don't work. Let me show you what works
Extras:
1. It has guidelines for more durable/recoverable workflows โ I've lost a lot of progress prior to these rules due to individual workflow agents failures
2. A guideline when to launch an agent or workflow and when to write a script โ ultracode fans know what I mean ๐
@gootecks@dzhng Because there are no triggers in the description, you can have a prompt in CLAUDE.md with instructions on when to use it until solution to this issue lands: https://t.co/FJ14zeo2nu