You will be able to read a first answer as a draft and work out what feedback it needs.
Open the brief you wrote for your AI assistant and put it next to the first draft it gave you before you decide what to do with that draft. Read on its own, a draft only tells you whether you like it. Read next to the brief, it shows you which parts to keep, which parts to change and whether your request left something out.
A first answer is rarely exactly right, even when the brief is strong. It is often close, though. The facts are present, the structure is roughly sensible and most of the sentences can stay. What it usually lacks is the final layer of judgement that comes from knowing the client and the history, and the model can only partly pick that up from your brief. Plan for one or two rounds of feedback as a normal part of the task, the same way you would mark up a first draft from a capable colleague.
This changes the question you ask of a first draft. Instead of asking whether it is right, ask how far it is from right and what would close the gap. Closing the gap usually takes a few targeted instructions.
Two common reactions skip giving feedback, which is the most effective step. One is to regenerate and hope for something better. The other is to rewrite the draft yourself. Regenerating without feedback means asking for a new attempt without telling the model what was wrong, and it is a gamble. The model samples its words, so each new attempt will differ (AI fundamentals lesson 3.3, "Why the same question gets a different answer", explains this). The new version may fix the problem, or fix it and introduce a new one, and you still haven't told the model what you wanted.
Marcus (an invented example) is an account manager at an IT support firm in Ubi. One of his oldest clients, a family-run trading company, has a contract coming up for renewal, and the fee is going up. He writes a good brief following module 1. It covers who the client is, how long they have worked together, the new fee and why it is rising. It also sets a 200-word limit and asks for a subject line plus body.
The draft comes back competent but stiff. It explains the price rise before saying anything about the relationship, and it ends by offering a call "at your earliest convenience", a phrase nobody at his firm would ever use with this client. Marcus's first instinct is to regenerate it and his second is to rewrite it himself. Both would skip the feedback step.
Reading against your brief means checking a draft part by part against the four parts of the brief that produced it. The four parts come from lesson 1.1, and each one gives you a check: right reader, purpose served, constraints followed and format used. When Marcus reads his draft this way, he finds:
Reader: wrong. It reads as if written to a stranger, so the relationship he described didn't come through strongly enough. Purpose: missed. He wanted the client to renew happily, but putting the fee before the relationship made it read like a bill. Constraints: met. It came in under 200 words. Format: met. It had a subject line and a body.
The check turns his vague feeling that the draft is "not quite right" into specific points he can act on. It also shows him which parts of his brief worked, so he can keep writing them the same way.
His feedback then says what to keep as well as what to change. If you leave out what to keep, a revision may rewrite the parts you liked along with the parts you didn't, and you swap one set of problems for another. Marcus tells the assistant to keep the paragraph about the new 24-hour helpdesk (also invented) word for word, and to keep the fee figure and the start date exactly as written. With the facts protected, the model is free to change the order and tone, which were the real problems. When you do this yourself, quote the exact parts, figures or dates you want kept, then name what should change. Lesson 4.2, "Feedback that gets a better second draft", covers how to write these instructions.
When a first draft is far off, the problem is usually in the brief rather than the draft. The signs are a draft aimed at the wrong reader, one that takes a stance you didn't want, or one that misses the point of the task. Feedback on the draft can't fix something that was never in the request. To test for this, ask yourself whether you can point to what was missing from the brief.
Suppose Marcus's draft had offered a discount he never mentioned. His brief would have needed the constraint "no discounts or special offers". Or suppose it had been addressed to the client's finance department instead of the owner he has known for nine years. Then his brief would have needed a clearer reader. In cases like these, fix the brief and run it again, and don't layer corrections on top of a bad draft. Lesson 4.3, "Edit, regenerate, or start a fresh chat", shows how to rerun it cleanly.
Fixing the brief has a second benefit. The improved brief is the one you save and reuse, so your next similar task starts from a better place.
A habit worth building is to read every first answer with three columns in mind: what to keep, what to change, and what your brief should have said. Find a recent first answer that is still in your chat history, along with the brief that produced it. Read it against the four parts of the brief and write at least one point in each column.
Take a first answer from an assistant and check it against your brief. List what to keep, what to change and what the brief left out.
Junxiong-WFG Organisation is an authorised representative of AIA Financial Advisers Private Limited (Reg. No. 201715016G).