principle-encode-lessons-in-structure

Encode the recurring correction as a lint rule, metadata flag, runtime check, or script instead of writing it again. If the rule genuinely requires human judgment, make the instruction more prominent and add an example of the failure mode.

Choose the strongest mechanism the situation allows: first an unrepresentable state that cannot compile, then a lint rule or banned API that fails CI, then a canonical helper, then a runtime check. Weaker guards become templates that agents copy.

Route one-off issues to a simple note, recurring fixes to a skill or lint rule, and systemic issues to a principle. The key decision is whether the correction is a one-off or a repeating pattern.

Textual instructions require readers to notice, remember, and comply, which makes them easy to miss. Structural mechanisms like lint rules and runtime checks enforce the rule without depending on cooperation.

Common anti-patterns include acknowledging feedback without recording it, recording it without routing it to implementation, and fixing one instance while leaving the recurring pattern intact. Each breaks the feedback loop.