Developer shares per-directory README practice to keep AI coding on track

At 23:14 on November 12, 2026, a developer published a practice on Juejin for keeping AI coding assistants from editing the wrong files. The approach is built around per-directory documentation: a README is placed in every layer of a project's directory tree, and each file records only the direct subdirectories and files that sit directly under it, effectively handing the model a map of which parts belong to whom. The stated rule of use is that AI tools must read the README for the target directory before making changes, and may trace upward through the directory structure no more than two levels. That constraint is meant to define the scope of edits and stop the model from browsing code without direction. On top of the documentation habit, the author adds a gap checker that can actually run, and wires it into CI so that mismatches between the README descriptions and the real directory structure are caught in the pipeline. The author claims the full setup, combining one README per directory, read-before-edit behavior, an upward look limited to two levels, and a working gap checker hooked into CI, delivers an exceptionally high return on effort. As of the available reporting, the practice has only been presented in this single Juejin post, with no follow-up developments or community responses documented in the material.