You’re already dangerous with these tools. Here’s what actually works.

The short version

Clipped names cost about 30% less per AI run and delivered 0 of 5 working features, twice. Full names delivered 2 of 5, then 4 of 5.

The numbers

• $0.44per run, clipped names
• $0.62per run, full names
• 0 and 0working features of 5, clipped names
• 2 and 4working features of 5, full names

The clipped version also read about 40% less code, 61% as much as the full-names version. Average costs were $0.4378 and $0.6246 a run.

In both clipped runs the pet list never showed up: run 1 rendered empty and run 2 returned an HTTP 500.

A cheap run is not a cheap feature

You have heard “keep it short, tokens cost money.” The cheaper run saved thirty percent and delivered nothing you could use. The full-names run cost a little more and got working features. Measure what ships.

How we ran it

One task, a pet-tracker app. Two runs per version, the same model. The full-names version also had its naming rules written into its instructions.

How the code was checked

A bigger model reviewed each run’s built code task by task against one rubric. It read the code and did not run the app.

Where the evidence lives

The reviewer’s notes and both branches are public in the aed-conventions repository under evidence/buildoff_2026-08-11. The naming rules behind the full-names version are in the AED book.