Simple prompts are working better for Nico lately: ask for something "super simple," or point the model at what it did last time.
People often hold back from AI because watching it do their job well and fast is intimidating; seeing it as a tool, not a replacement, makes them more willing to try.
To start adoption, show someone AI doing a process task they hate that eats up their time.
AI's return is time saved, which makes it an organization-level question: what the saved time goes to.
Instead of an impact-versus-effort grid, fix one small, visible, everyday irritation for many people, and let that win build trust for bigger projects.
Keep the core instruction file short, with modular files around it; Nico's points to everything else in about 25 or 30 lines.
Give each project its own memory, and set a rule that nothing counts as done until it has been tested.
Hooks fire before or after an agent acts; loops keep going and burn tokens, like a HubSpot workflow with re-enrollment.
Check finished work with a separate test that sees only the original request and what was delivered.
Models estimate time the way a human would; Nico keeps a record of how long tasks actually take, and works backward from the result he wants.
