Building with AI
I helped teams at Superteam's Blitz hackathon plan a small working application, use AI during the build, and prepare to put it online.
Superteam · Singapore · 14 March 2026
Choosing What to Build
With a day to turn an idea into a first version, teams need to choose what the application must actually do. I used the workshop to distinguish between asking AI for an answer, letting it carry out steps, and connecting it to an existing service. That choice helps define a build the team can finish.
The session followed planning, building, and getting the application online, keeping that final step in scope from the beginning. I also covered how to give the assistant relevant information and check its code or text, since confident output can still be wrong or out of date.
Scoping an AI Build for a Hackathon
Model, Harness, and Application
I organized the fundamentals around prediction, autonomy, and integration. The model generates an answer, the harness manages tools and execution, and the application wrapper gives users a way to work with the result. Distinguishing these layers helps a team decide whether its idea needs a better answer, a sequence of actions, or a connection to an existing service.
Task Context and Output Review
Knowledge cutoff, context management, and hallucination affect how a team supplies information and checks generated code or text. Relevant task context gives the assistant something concrete to work from, but does not remove the need to review its output. The workshop kept planning, building, and getting the application online in the same scope so a first version included a usable delivery path.