In the last post we generated 1000 architect personas. What I started to notice when we created designs is we’d really skewed the output with our example prompting.
Lots of Zero Trust, Security architects. I could go back the randomly generate a set of titles and backgrounds myself but I really wanted to make the as much AI generated as possible
Luckily, reasoning models could come to our rescue and seemed like they might be a better fit.
“A reasoning model is like a senior co-worker – you can give them a goal to achieve and trust them to work out the details”.
It meant I had to review our prompting as the guidance for reasoning models says:
- Keep prompts simple and direct
- Try zero-short first
So we modified it to be a lot more generic and just pass the items we want to force in the code, removing the specific json example:
What that gave us was a better range or levels in the cohort:
