Eval Loop
Iterative quality diagnosis that runs targeted fix-and-check cycles until output meets a defined quality threshold or hits max iterations.
Eval Loop is the skill for "this is almost right but not quite." It takes content or code, evaluates it against explicit criteria, identifies the top 3 issues, fixes them, re-evaluates, and repeats until the threshold is met or a maximum iteration count is reached. Each iteration is scoped: fix only the flagged issues, do not touch anything else. This prevents the common AI failure mode where "improving" correct sections while fixing broken ones introduces new regressions. The loop maintains an iteration log showing what changed at each step and why. Typical convergence: 2-4 iterations for content, 3-6 for code. The escape hatch fires if the quality score plateaus across 2 consecutive iterations — continuing past that point burns tokens without progress.
Where it shows up:
Related Terms
Methodology
Quality Gate
An automated checkpoint in a skill chain that evaluates output against defined criteria before allow...
Methodology
Iteration Log
An append-only record of actions taken during PRD execution, providing an audit trail of what was do...
Methodology
Escape Hatch
Predefined conditions in a PRD that trigger automated execution to stop and request human input inst...