The original title is: "How AI Agent Skills Degrade Under Agent Load — and Three Properties That Prevent It"

Short summary

This article examines how AI coding skills built for human use degrade predictably under agent load — producing output that looks correct but is silently wrong. Three properties determine survival: idempotency, output stability, and failure specificity. The author demonstrates this by running a Gherkin quality evaluator five times with identical input but different framings, showing how HTTP status codes and assumptions drift based on prompt phrasing alone.

• •Skills built for humans degrade under agent load in predictable ways
• •Three survival properties: idempotency, output stability, failure specificity
• •Demonstrated with a Gherkin evaluator producing different results across five identical-input runs

Generated with AI, which can make mistakes.