Quality ratchet

A quality ratchet is a practice of converting recurring failures into durable checks, examples, or instructions, then rerunning those checks after later changes.

A quality ratchet converts a discovered failure into a durable check. If an agent writes an unsupported claim in a weekly report, you correct the report, add a sourced-claim test or review criterion, and rerun that check on later reports. The new check makes the same failure easier to catch; it does not guarantee that output quality will rise forever or that the failure can never return. Models, inputs, tools, and instructions still change, so the ratchet works only when its checks keep running and someone repairs regressions.

Builder example

A correction confined to one chat disappears with that chat. A regression case stored beside the workflow can survive prompt rewrites and model changes. Keep the failing example, the expected behavior, and the check together, then run them before releasing a revised workflow. This turns a lesson from one bad output into evidence that later versions still meet the standard.

Common confusion: A quality ratchet is not a promise of monotonic improvement and it is not an ever-growing prompt. A useful ratchet represents failures with the smallest durable mechanism that can detect them, such as a test case, schema check, rubric item, or short instruction. Remove checks that no longer represent the intended behavior and consolidate duplicates.