Steering the average

What you'll walk away with: the two moves that decide what comes out — pushing the average off center on purpose, and counting the rules instead of trusting them.

You need: both tabs, about 20 minutes, and one real thing you have to write this week that you could forward outside the firm — an email, a summary, a plan. The duel runs in your browser.

Every lesson so far has been about what goes into the window. This one is about what comes out, and they are the same question: the reply is the continuation the frozen odds score highest over the window you built. So there are only ever two jobs — get the right things in, and push the result off the middle. Everything filed under prompt engineering is one of those two jobs wearing a name. This lesson is the second job: leave the window flabby and the reply lands on the average of what you loaded; load it so the average sits where you want it.

The pull of the average

You wrote a sharp request — your voice, your angle, a real opinion — and the reply reads like everyone's newsletter. No malfunction. Bland output is the mechanism working: the machine returns the statistically expected continuation of the window you built, and the statistically expected continuation of most professional writing is hedged, balanced, and shaped like the middle of its genre. Your one distinctive sentence was outvoted by everything else in the pile. The shaping you read about in lesson 02 pushes the same direction — the example answers it learned from were written to be helpful, balanced, and complete, and that polish lives near the middle of the road. Two currents, one drift: toward the average.

The move: move the average on purpose. Name the audience. Force the format. Ban the hedge you don't want — an instruction like end with one recommendation, not a balance of considerations is window text voting against the genre's middle. Paste an example of the register you want — an example in the window is the strongest vote you can cast against habits from training.

Act 1 · THE DUEL opens first: press Run both and read the two replies side by side — same raw update, same machine, two prompts — then press Send A or Send B for the one you would actually send. WHAT MOVED THE AVERAGE names the four votes prompt B cast. No live model runs; the footnote says SIMULATED. Press 2 · THE LOOP for the same craft made structural — fresh windows judging each other's work — the cheap form lesson 05 gave you; the lab Red-team critical work runs it for real.

Too many instructions

Now the same mechanism from the instructions side — because the obvious fix for everything above is "add more rules," and that has a failure shape of its own. Lesson 07 held one rule fixed while the pile grew; here the window stays short and the rules multiply.

You watched this in lesson 02 without naming it. The x-ray's instructions block — still there to check — said if the user asks for something the document cannot support, say what is missing, and the memo-shaped summary won the draw anyway. Instructions tilt the odds; they don't command.

So stack constraints and watch what tilting does. Every rule you add narrows what counts as a good continuation. Add enough at once and no continuation scores well on all of them — and nothing forces the machine to stop, error, or ask which rule to break. It writes the best-scoring continuation available and quietly sacrifices the rules that fit worst. There is no error signal for a violated instruction. The reply arrives fluent, confident, and missing constraint number eleven.

There is no rate to quote here: the count is yours to take, in your own tab, in two minutes. Number the rules before you send; tally the reply after. When a rule does not survive, nothing tells you — so you are the thing that tells you.

Two moves:

• Fewer constraints, ranked. Decide which rules are load-bearing and say so — the figure must come from the attached sheet outranks keep the tone warm, and the window should know it.
• Split the passes. Draft under the few rules that matter most. Revise under the next few. Check the survivors each pass — a dropped rule you catch in round two cost you a minute, not the deliverable.

This is the instructions-side twin of the balance in lesson 07's restart — enough of your material that it isn't guessing, little enough that nothing drowns; here, enough instruction that it aims true, few enough that none get dropped.

Do this in ChatGPT and Claude

One message you actually owe someone this week — an email, a summary, a plan — and a fresh chat for every send, in both tabs: a window with nothing of yours in it.

Round one — the duel, live. Predict, in one word, written down, what you expect back.

Write a short email to my team about [the thing you actually need to tell them this week].

Then a new chat, not a follow-up — the first draft in the window would vote for itself:

Write a short email to my team about [the same thing]. You are writing for someone with thirty seconds. Three sentences. Open with the decision, not the background. Name one next step with a date on it. No hedging words — no "potentially," no "approximately," no "generally."

Observe: put the two replies side by side and mark every hedge, every balancing clause, every line of background the second one does not contain. Explain: the same frozen odds ran twice; only the window changed — every clause you added voted against the middle of the genre. The point is the gap inside one tab, not between products.

Round two — twelve rules, and a count. Fresh chat, both tabs. Predict on paper: which numbers come back missing?

Write a short email to my team about [the same thing], and follow all twelve rules.

1. Exactly three sentences.
2. Open with the decision, not the background.
3. Name one next step with a date on it.
4. No hedging words — no "potentially," "approximately," or "generally."
5. No greeting line.
6. No sign-off line.
7. Include exactly one number.
8. Do not use the word "team."
9. Keep every sentence under twenty words.
10. No em dashes.
11. No bullet points.
12. End with a question.

Do not ask it whether it followed the rules. A reply about a reply is another likely continuation over a window that now holds the draft it is grading — not an audit. Tick the twelve yourself; every one is checkable from the reply alone — count the sentences, scan for the banned words and the em dashes, look for the date, the single number, the question mark, and read the first sentence for the decision. Write down the numbers that did not survive. If all twelve did, that is a real measurement at this count; add three more of the same kind and send again — then re-place from that round.

Then re-place them — the whole countermeasure: one more fresh chat, the reply pasted in, and only the rules that went missing.

Here is a draft:

[paste the reply you got]

Revise it so that it obeys these rules, and change nothing else: [paste only the numbered rules that did not survive].

Observe: count again. Explain: a dropped rule is a shape with a countermeasure, not a limit of the machine — and any gap between the numbers you guessed and the ones that went missing is the argument for counting rather than trusting. Tally each tab against the same twelve, then compare: what repeats is the shape to plan around; what differs is one product on one day, never a ranking.

---

Check your understanding: a draft comes back bland, and a second draft written under twelve rules comes back missing two of them. What is the one mechanism under both, and what do you do about each?

Check yourself first — then open one good answer+−

One draw under both: the reply is the continuation the odds score highest over the window you built, and the window decides where "highest" sits. A flabby ask scores highest on the middle of its genre, so you move it by voting — an audience, a format, a banned hedge, a pasted example of the register. Twelve rules at once and no continuation scores well on all of them, so one is left behind with no error line. You do the tracking: number the rules, tally the reply, rank the load-bearing ones, and re-place the missing ones in a second pass.

Your answer should have hit one machine and two moves: push the average off center on purpose, and count the rules instead of trusting them.

Use this the moment a draft comes back generic. Before you re-send, audit what else you loaded that votes for generic — the boilerplate brief, the safe template, the average example — then cast louder votes against it. The output is only ever as good as what you put in front of it, and this is how you aim it. Next: Hallucination, and the word for absence — what it does when the window has no answer in it.