“One more time. Half as loud, twice as pretty.”
Nobody can tell you who said it first, but every horn player alive has heard some version of it from a bandleader mid-rehearsal. It survives without an author because it is a rule of the house rather than a quote: compressed craft, passed mouth to mouth, because it fixes something that goes wrong in every ensemble, in every era.
Given the choice, players get louder.
Volume feels like contribution. Volume feels like presence. And volume is the cheapest thing a musician can produce, costing no technique, no listening, and no restraint. Which is exactly why the direction exists. The bandleader is asking for more music and less noise, and experience has taught them that the knob for one is usually the knob for the other, turned the opposite way.
What actually happens when a section plays at half volume is that everyone can suddenly hear everyone. Tuning problems that loudness was hiding become audible, and get fixed, because a chord out of tune at pianissimo is obvious in a way it never is at full volume. Phrasing has room to exist, since a line you are shouting cannot be shaped. The ensemble blend appears. The soloist comes through. The music that was always in the chart, buried under everyone’s enthusiasm, stands up, in tempo and in tune.
The models got loud.
We work in AI, which is the loudest industry on earth, and we want this rule stitched into every system prompt and every product review.
When you ask a simple question, you receive an essay. Request a summary and get an introduction, three sections with headers, a bulleted recap, and a closing paragraph restating the opening one. The industry measures the shouting as though it were the product: tokens generated, words per second, and output length as a benchmark. Volume as contribution, volume as presence, the cheapest thing a model can produce, celebrated on the dashboard.
The bandleader’s trade is sitting right there. The knob for more music is the volume knob, turned down. The answer that says the one thing that matters and stops. The report whose every sentence survived because somebody, or something, asked whether it earned its place. The system that answers a yes-or-no question with the word yes, the reason, and an ending.
Half as Loud.
When output gets quieter, the same acoustics arrive that a band discovers at half volume: errors that verbosity was hiding become visible, structure appears because an answer you are shouting cannot be shaped, and the reader (who is the singer in this arrangement) finally comes through, because the work was written to be heard.
We built our house on this rule, so here is the specific practice.
Every substantive answer our systems return leads with what we call a whisper, meaning the condensed verdict, up top, half as loud as the full analysis that follows it and (we hope) twice as pretty. The long version is there for whoever needs it. The whisper is the performance while everything after it is the appendix. It changed how we write, and then it changed how we think, because writing a good whisper requires having finished the analysis.
The player who can play it half as loud is the one who actually knows the part. Loud is where uncertainty hides. One more time. Half as loud. Twice as pretty.