Just wanted to share that using GPT-5 Pro to rewri...
# success-stories
r
Just wanted to share that using GPT-5 Pro to rewrite my prompts on my pipelines for Openhands + GPT-5 has made a ton of difference. if you are running something custom on top of openhands and you have a very prompt heavy solution. be sure to try using the prompting guide + your current prompt and getting gpt-5 pro to rewrite it.
🙌 4
a
Are you switching to GPT5 as your new standard?
r
Experimenting it... I fully switched for now, but depending on users reports I might go back.
so far the main complain is that it is now taking a long long time compared to what it took before.
(and before the prompt change that the result was subpar... it seems better now but I wait for the reports)
without tool calling enabled: failed most things. With tool calling enabled: failed some things. With new prompt: everything seems to work.
a
Makes sense, thanks My main issue is thinking we will have to rewrite every microagent anytime we make an LLM Switch
r
Yeah, but I only plan to switch when it is worthy. So far nothing beats claude. now it is different, even with everything against it , gpt-5 beat claude... and people are figuring out small changes that even improve the result.
đź’Ż 1
so in that case it might be worthy..
a
Yea, we will see if Anthropic drops the price of Claude 4 in a reaction move
r
I am curious to see what would be the result of using sonnet with this gpt-5 prompt... I’m curious how Sonnet will perform once we adapt the prompts for GPT-5’s style. Will it match GPT-5’s score on what were originally “Sonnet-optimised” tasks? And when the tables turn, will reviewers stay as unforgiving toward non-GPT-5 models?
The model that impressed me the most so far was kimi k2.... non reasoning and fighting against them. I imagine what a reasoning kimi k2 with prompts engineered to it would be capable.
c
FWIW I see some reports in the wild, and so far it seems to match my experience, that GPT-5 is more prompt-sensitive than Sonnet 4. e.g. Sonnet dealt a bit better with ambiguity, but if people tell GPT-5 what to do, it does it
r
100% , I had some instructions repeated in different flavors to convince sonnet 4. With gpt-5 I removed the bulk of it.
pretty sure if I run sonnet 4 with this prompt again I will have all the issues that I used to have with previous claude versions. GPT-5 seems overall much much smarter (and steerable)
👍 1
a
DSPy is intended to create an abstraction layer between the LLM + prompt and the desired behavior, so that you can more easily switch LLMs
👍 1