Notes from the workshop.

Half of what we scope does not need a model. The test we use: agents only where open-ended probing earns its cost, ordinary code everywhere else.

Eight thousand words of playbook fit in a cached prompt. Why we shipped a constant instead of retrieval, and the honest line where that decision flips.

Mexican banks offer no Plaid-style sync, so the data arrives as a PDF. Here is how we split that job between a model and ordinary code, and why an import that refuses beats one that is quietly wrong.

A chatbot answers questions. An agent takes the next step for you. Here is what that difference looks like on WhatsApp, and why the channel makes it real.

The fastest way to miss a WhatsApp launch date is to discover Meta Business verification in the last week. Here is what it checks, why it stalls timelines, and how to de-risk it.

Most AI work is sold by the hour and ends without a finished product. We sell a plan and a price, agreed before we start. Here is how fixed scope works, and where it does not.

The demo is the easy part. In week one we ship a working slice instead, one real path running end to end. Here is why that de-risks the whole build.