Hi everyone,
I’m noticing a big difference in behavior between a custom GPT and an OpenAI Assistant — even though I’m using exactly the same instructions and files in both.
The GPT (even using gpt-4o-mini) follows the instructions well and produces responses that align with what I expect. However, when I test the same setup using the Assistant (via Playground), the results are very different — even after trying different models and low temperatures (0 to 0.2).
One major issue is message splitting. I’ve been very explicit in my system instructions, repeatedly stating things like:
“Never delay or split your response. Always send full replies (greeting, policy, and content) as one single message.”
and:
“Always respond with a single, complete message — never divide into multiple parts. Do NOT break the response into separate replies under any circumstances — including the first interaction.”
Despite this, the Assistant continues to split replies, especially on the first interaction — which breaks the expected flow, particularly for WhatsApp-based use.
Has anyone else run into this? Is there something I’m missing about how the Assistant handles instructions compared to a custom GPT?
Any help would be much appreciated!