I have Custom Instructions that can get ignored in a chat.
If I want control over the outcome or am doing anything remotely complex, I make a GPT and provide knowledge files, and if there is an API I want to use and it’s huge, I will chop it down with Swagger Editor or another custom GPT (grab the GET operations…) and make Actions.
This leads me to chaining agents with a specialty; the third party API, the general requirement, the first-party API, and code generators with knowledge for documentation and example code.
I chain these together with @ and go directly to town with run, eval, enhance, check-in loops.
I have turned out MVPs in multiple languages for a bake-off in the time it might have taken to select the first toolkit for evaluation. We’re running boiler plate example code tweaked to purpose. With 4o, the memory and consistency is really improved. It’s not a full rewrite every time, it’s honoring atomic requests.
If I want control over the outcome or am doing anything remotely complex, I make a GPT and provide knowledge files, and if there is an API I want to use and it’s huge, I will chop it down with Swagger Editor or another custom GPT (grab the GET operations…) and make Actions.
This leads me to chaining agents with a specialty; the third party API, the general requirement, the first-party API, and code generators with knowledge for documentation and example code.
I chain these together with @ and go directly to town with run, eval, enhance, check-in loops.
I have turned out MVPs in multiple languages for a bake-off in the time it might have taken to select the first toolkit for evaluation. We’re running boiler plate example code tweaked to purpose. With 4o, the memory and consistency is really improved. It’s not a full rewrite every time, it’s honoring atomic requests.