Main idea

The instruction file of competitor_analyze grew from 18,000 to 27,700 bytes in three weeks, swollen by format templates written directly inside it. Every addition is settled on every call, whatever the work requested: five thousand extra tokens, a hundred runs a month, half a million tokens spent without any task having required it.

A prompt grows without resistance because each added line looks free at the moment you write it. It is the accumulation that gets paid for, and it appears in no step of the workflow. Externalizing heavy templates and loading them where they serve turns a fixed cost into a conditional one.

Why it matters

This makes the prompt an architectural expense rather than an exercise in writing: its length is a design choice, just like what you decide to load from a database.

It also explains a counter-intuitive effect — a skill can cost more after being improved, because it was documented inside its own instructions.

Nuances and limits

Fixed context is cached after the first turn: in steady state it weighs only a fraction of its volume. The inflation therefore matters mostly for skills relaunched cold, and the gap deserves to be flagged as such in a report.

And a slimmed-down prompt that causes round trips or formatting errors costs more than it saves.

Open questions

• How do you know which part of an instruction file is actually mobilized during a given run?