AI models generate code well from examples, but repeatedly supplying large amounts of context can be costly and may reduce consistency. This talk explores whether code templating tools can act as reusable skills: capturing the essential structure of an example while using fewer tokens to generate similar implementations.

We’ll test the idea at three scales: small files, a medium-sized feature, and project scaffolding - comparing template-driven generation with a conventional context-heavy approach. The result is intentionally uncertain and open-ended: the experiment may reveal token savings and more reliable output, or show where templates become too rigid or costly to maintain.

Flutter Cracow Meetup, Kraków 2026