agent-legible-codebase
Installation
SKILL.md
The Legible Codebase
Two engineers run the same agent on the same model against the same prompt and report wildly different experiences. One is editing a repo with sharp tests and clean modules. The other is editing a repo whose most-touched file is a 2,000-line controller with three TODOs from 2022. The variance the second engineer is fighting is repo variance dressed up as model variance.
Instruction is a claim. Imitation is evidence. Feedback is consequence. Claims without evidence read as marketing. When general guidance and a specific local example conflict, the example wins — it answers the agent's actual question.
This is why "write more rules" is almost never the right first move.
Start here: the three-question diagnostic
Hand the agent a representative task. Read the diff.