A draft can look clean and still be wrong in ways you only notice on the fortieth one. Taste does not scale.

This book turns output quality into something you can measure, gate on, and track across thousands of generations. It builds the measurement layer for Claude Code operators who produce text at volume: a hard-fail gate that blocks bad output before it ships, a quality spec mined from the failures you actually hit rather than a list of best practices, sentence-level metrics that put a number on the things people usually judge by feel, and a drift report that shows where Claude's output wanders once you look at the aggregate.

Seven chapters, seven artifacts.

Chapter 1 makes the case for a number over a verdict and defines the measurable axes of text quality.

Chapter 2 builds the hard-fail gate - what belongs in a blocking check versus a warning, and a scanner that runs on every output.

Chapter 3 is the method that sets this book apart: building the quality spec from observed failures, turning each real miss into one measurable rule.

Chapter 4 quantifies the hard-to-measure - opener diversity, repetition, negation density, vocabulary spread - and insists on showing the count before calling anything clean.

Chapter 5 measures instruction adherence across many outputs and flags drift from the spec.

Chapter 6 reads the failure log in aggregate, weighting by severity and recency to surface where quality breaks down.

Chapter 7 assembles the whole thing into one reproducible pass that gates a batch and signs it off with an acceptance certificate.

The device-level craft of removing specific AI tells is its own book - *Removing AI Tells* (Book 6). This one builds the system that measures and enforces quality at scale. By the last chapter you can hand Claude a quality bar it cannot quietly miss.