REFERENCE 10 / 14 · CORE CONCEPT · 5 MIN
Steering the average
Output drifts toward the average of what you loaded. Prompting is how you push that average off center.
Whatever you load, the output drifts toward its average. That's not a flaw — it's the whole mechanism. The machine gives you the statistically expected continuation of the window you built. So the craft has one name: move the average on purpose.
The duel above shows the everyday version. Prompt A got the average summary — the one everyone gets, hedged and shapeless. Prompt B named an audience, forced a format, banned the hedging, and demanded a decision. Same machine, same facts, different average.
Three levers, in increasing order of power:
- Words in the prompt. Audience, format, register, constraints. Cheap and immediate — this is what "prompt engineering" mostly is.
- Borrowed averages. A persona ("you are a panel of Jocko, Buffett, and Hickey…") isn't theater — it pulls the output toward the register and concerns of specific bodies of writing. You're choosing which averages to drift toward.
- Structure. The loop in act two doesn't ask the model to be better; it arranges fresh windows so problems get surfaced. A fresh window removes the built-up commitments of the first one — but it is the same model with the same instincts, so agreement between runs is comfort, not proof. What converts the critiques into trust is the other half: resolve each one against a source of truth, a recomputation, or your own read. De-anchoring finds the challenges; the source settles them.
Which is also the honest word on trust: a fluent answer is the average sounding like itself, not evidence of being right. You don't fix that with vibes; you fix it with structure — sources of truth and fresh-window checks.
Say it back in one line: why does a critic in a fresh window find problems the original window's critic misses?