How a prompted-first product usually goes: the happy path works and the demo looks great. Then someone asks about roles, permissions, or what happens when data is missing, and nobody has thought about it. Every fix after that is bolted onto code that wasn't designed for it.
#AI #Startups #LLM #MVP
How a prompted-first product usually goes: the happy path works and the demo looks great. Then someone asks about roles, permissions, or what happens when data is missing, and nobody has thought about it. Every fix after that is bolted onto code that wasn't designed for it. #AI #Startups #LLM #MVP
0 vue
Publié le 19/11/2026 14:00
How a prompted-first product usually goes: the happy path works and the demo looks great. Then someone asks about roles, permissions, or what happens when data is missing, and nobody has thought about it. Every fix after that is bolted onto code that wasn't designed for it.
#AI #Startups #LLM #MVP
1 commentaire
?
That isn't the model's fault. It built what it was told. So I write the spec first: a plain-language brief from discovery, then deep research turns it into roles, flows, data model, integrations, edge cases, security and acceptance criteria, with a source behind each assumption.
Then version one gets locked. Not frozen: a new idea becomes a decision (update the spec on purpose, or park it for v2) instead of drift. BlastRocket was a written spec before it was software, and choices like failing closed on missing evidence came from that, not from the code.