Iterating: the first answer is a draft

1What a follow-up must say to be obeyed

Microsoft writes that the first answer is a draft and recommends continuing the conversation: "make it shorter", "use a more formal tone", "add a column on risks". Each turn keeps the context of the previous turns. So you do not have to repeat the whole prompt: you adjust.

Iterating is steering. The word iteration names this cycle: read the answer, judge it against your goal, formulate a correction, follow up. The skill at play is called steering: redirecting the model without losing what was already good.

An effective follow-up has three parts. What to keep: "keep the structure and the first paragraph". What to change, in checkable terms: "80 words at most", "start with the key figure". And, if possible, an example of the expected result. "That's not it" teaches the model nothing; "the second bullet is off topic, tighten it around the budget" puts it back on track.

Two reflexes complete the method. Change only one thing per turn: if you alter the tone, the length and the format at once, you will not know what produced the improvement. And ask Copilot for its assumptions before it produces anything: "list what you are assuming about my request, then answer". You correct the wrong assumptions before it builds on them.

One reservation, finally. Asking Copilot to "check its answer" without giving it a source rarely improves the facts: research shows that models correct themselves poorly on reasoning on their own. The useful critique comes from you, or from a confrontation with a provided document.

Tap to enlarge

2An expenses announcement shortened without loss

An internal communications officer wants to announce a new expense claim procedure.

Turn 1 (Initial framing prompt and flaw identification).

Draft an internal memo to announce the new expense claim procedure taking effect on 1 November.

Observed response: Copilot produces a verbose 240-word draft, heavily loaded with technical jargon ("multi-tier approval workflow"), burying the critical deadline at the very end.

Turn 2 (Surgical calibration follow-up).

Keep the three-paragraph structure and the 1 November date. Rewrite in 120 words at most for colleagues who have never used the tool. Replace every technical term with an explanation in brackets. Start with what everyone must do from their very first expense claim.

Observed response: Three clear paragraphs totalling 110 words. The date is preserved, "approval workflow" became "approval chain (your manager, then accounting)", and the immediate required action sits on the first line.

What changes. The follow-up names what to keep, so nothing important disappears. The corrections are numbered and checkable: you can count the words. The recipient is restated, which guides the choice of vocabulary.

3Rewrite an invoice follow-up in three parts

You asked Copilot Chat for a reminder email to a client who has not paid a 3,400 euro invoice that fell due twelve days ago. The first draft runs to 220 words, addresses the client too casually and mentions penalties that your contract does not provide for. Here is the starting follow-up, to be improved:

"No, that's no good, make it more professional."

Rewrite this follow-up in three parts: what to keep (the invoice number, the amount, the request for payment within eight days), what to change (formal address, 90 words at most, no mention of penalties) and an example opening line. Run it, then do a second turn that changes only one thing. The composer below helps you word the initial prompt; the follow-up is written directly in the conversation.

Self-assessment grid: (a) every follow-up says what to keep and what to change; (b) only one thing changes per turn; (c) the final email fits in 90 words and mentions no penalty.

Open the prompt composer