1. Find where people wait
Most of my products started when I saw someone waiting on someone else. At BharatPe, engineers waited 14 days for a database query.
2. Write down what not to build
Every spec I write has a “Not built” section. For Atlas, version one had no automatic fixes: it explains, and a person decides. A non-goal written on day one prevents a month of debate.
3. Put the rules in code
A rule in a document gets broken. In Sakhi, safety checks run before and after the model on every message.
4. Build the first version myself
I build first versions with AI coding agents, in parallel streams, and I review what comes back. A working version answers what a spec cannot: does the user understand it, and does the model behave?
5. Test whether it feels right
In Sakhi, every automated test passed, and the agent still sounded like a phone menu. Now I ask of every test suite: would it fail if a real user found the product wrong? The full story: why our health agent sounded scripted.
6. Decide with evidence
I write each major decision as a one-page explainer: conclusion first, evidence next, open questions last. Everyone argues about the same facts.