All field notes

AI engineering6 min read

AI-readable frontends use fewer guesses, not simply fewer tokens

Token efficiency comes from stable names, finite choices, examples that compile, and machine-readable boundaries—not from removing the context an agent needs.

01

Ambiguity is expensive

When an agent must search source files, infer naming conventions, invent component APIs, and repair unsupported classes, each guess consumes context and creates another opportunity for drift.

A compact registry with literal import paths, supported variants, examples, status, and constraints gives the agent a bounded decision surface.

02

Compression should preserve decisions

The goal is not the shortest possible prompt. The goal is the smallest complete contract. Stable identifiers can replace repeated prose, while examples can demonstrate exact syntax more efficiently than a long explanation.

  • Publish exact utility and component names.
  • Expose compatibility and stability status.
  • Provide one verified example per common task.
  • Return actionable diagnostics when a choice is invalid.
03

Measure successful work, not prompt length

A shorter prompt that produces three repair cycles is not cheaper. Evaluate total tokens, tool calls, elapsed time, build failures, and human correction needed to reach an accepted result.

AI-readable infrastructure pays off when the first implementation is closer to the current product contract and easier to verify automatically.