ACI-024: The Prompt That Worked

Thesis

The quality of a task definition determines the quality of agentic output more than any other variable. Good task definitions constrain scope, declare dependencies, and validate architecture before committing to a full build. Bad task definitions delegate outcomes without constraints, and agents fill the gap with scope expansion, missing dependencies, and local optimization.

Story

Laksa is a local-first CMS built with Elixir/Phoenix/Ash over 4 months. It started as a port of an existing Rails application (Postcard) and evolved into something larger. Across 48 completed steel threads, the plan-vs-outcome divergence followed a consistent pattern: task definitions that described what to build without constraining how much to build produced scope explosions. Task definitions that asked a question before committing to work produced clean outcomes.