Stop Shipping AI Slop

Three pieces from one session. The session started with an X post arguing that AI slop is an output-side problem — not a prompt problem — and ended with a running quality gate, a build walkthrough with verbatim scores, and a meta-analysis of the six decisions that made the gate small enough to actually ship.

Read what matches the problem you’re working on. The pieces don’t need to be read in sequence.

Contents

Article	What it covers
The Scoreboard Is the Moat	Why better prompts don’t stop slop, what a measurement layer looks like, and the silent surface (product output) that costs more than the loud one
I Stopped Trying to Prompt My Way Out of AI Slop. I Built a Scoreboard Instead.	The full build session: the 0.85 and the 0.16, the blocked regression, the production dip that rewrote the test suite, and why the gate never edits the work
Every Decision Was a Decision to Build Less	The reasoning pattern under the build — six questions that each removed work before any code got written, formalized as the Subtraction Pass

About the session

Origin: June 1, 2026. One build session: a quality eval loop ported from a platform (Hermes) onto a folder-based ambient intelligence architecture — run live, with real scores, real blocks, and a compounding test suite.

The three pieces come from the same session but approach it from different angles. The first is the strategic argument (why measurement matters). The second is the build diary (what happened). The third is the meta-layer (how the decisions got made). Together they cover the what, the how, and the why-it-stays-small.