There is never "one" prompt per say.. it's usually a workflow or something with the right set of information in the context and right set of tools.
One thing I'm proud of is that at my work (the product is a child facing toy / bot), i created an agent loop that tests our AI layer with adversarial questions (that it should block) based on trending non child friendly topics. This was done manually before, and so my agent loop saves a lot of time
Hard agree with this My fave workflow I did was making the first fully regulatory approved LLM workflow for a global retail bank, using GPT 3.5 in June 2023 on Azure early access. That stack is still the reference for a lot of the work we've done since and currently.
One thing I'm proud of is that at my work (the product is a child facing toy / bot), i created an agent loop that tests our AI layer with adversarial questions (that it should block) based on trending non child friendly topics. This was done manually before, and so my agent loop saves a lot of time
Works every time.