We believe context is the key to unlocking AI's potential

Dex Horthy argues that AI coding tools fail in large codebases not because models aren't smart enough, but because context becomes too big for them to handle.

This applies to the prototypes, and the design systems we build.

He introduces frequent intentional compaction — periodically summarizing and refining the context fed to AI instead of letting it grow unbounded. Using this approach, his team shipped week-long Rust projects (100K+ LOC) in a single day with Claude Code, passing expert review. Referencing a Stanford study showing AI can reduce productivity in complex environments, Horthy reinforces that productivity gains require disciplined workflows — not better models. Structure and intention beat vibes.

Our take: We have also seen that AI tools struggle with large, complex design systems and/or prototypes. This is a fundamental limitation of current models and agents. We have seen many vested interests argue for the opposite — that the problem is just "not using the tools right" or "not prompting well enough". We dismiss all of these arguments and strongly believe this will be the case for the next 5-10 years, more likely 10 years than 5 years. The only way to get better results from AI tools is to build systems and workflows that work with the limitations of context size.

Context Engineering Talk by Dex Horthy