Refine Claude's First Answer Into the Right One
Getting Started with Claude introduced the basic move: treat the first answer as a draft, react to it, ask for another pass. What it doesn't cover is what to actually say in that reaction. "Make it better" or "try again" gives Claude almost nothing to work with, which is why those follow-ups so often come back barely different from the first attempt.
"Try again" and "fix the second paragraph, it's too formal" ask for the same thing. Only one of them tells Claude what changed.
Name the category, not just the feeling
A first draft that isn't right is usually wrong in one of a small number of ways. Naming which one turns a vague reaction into a specific instruction.
Wrong tone. The content is right but it reads too stiff, too casual, too apologetic, or too aggressive for who's going to read it.
Wrong scope. It covers too much, too little, or the wrong part of the request entirely, common when a prompt asked for several things at once.
Missing detail. A specific fact, number, or example needed to make the draft usable isn't in there, usually because you didn't include it in the first prompt either.
Wrong assumption. Claude guessed at something you didn't specify, and guessed wrong: it assumed a different audience, a different deadline, a different starting point than the real one.
Wrong format. The content is close but the shape is off: a paragraph where you needed a list, three options where you needed one, a long version where you needed a short one.
Tip
Before reacting to a draft, silently sort what's wrong with it into one of these five buckets. The bucket you land on tells you almost exactly what to say back.
What each correction actually sounds like
Wrong tone: "Same content, but less formal, like you're talking to a teammate, not a client."
Wrong scope: "Cut the third paragraph entirely, it's answering a question I didn't ask. Keep the rest as is."
Missing detail: "Add that this only affects the East Coast team, that got left out."
Wrong assumption: "This assumes the reader already knows the project context. They don't, add one sentence of background first."
Wrong format: "Turn this into three bullet points instead of a paragraph, one per option."
Each of those points at one thing and leaves the rest of the draft untouched. That's the difference between a correction and a restart: a correction tells Claude exactly what to change and what to keep.
This is close, but it's addressed to the whole company. It should sound like it's just for our five-person team, people who already know the backstory. Keep the structure, loosen the tone, and cut the second paragraph, which re-explains context they already have.
”
Inside Claude Tutorial
Pointing at a specific problem beats a vague reaction everywhere.
This isn't just a writing habit. The app's lessons apply the same targeted-feedback skill to reviewing code, evaluating a plan, and giving feedback to other people, not only revising Claude's drafts.
Refining across more than one round
A single correction is often enough, but not always. When it takes two or three rounds, the same rule applies each time: react to what's actually in front of you, not to your memory of the first draft. A second-round correction that says "also fix the ending" while ignoring what changed in round one usually means the two instructions quietly conflict, and Claude has to guess which one wins.
A useful habit for a longer refinement: after two or three rounds, read the current draft as if you'd never seen the earlier versions. That's the version a reader will actually see, and it's easy to lose track of what's still wrong once you've been staring at three drafts in a row.
The mistake that throws away all of this
Common mistake
Starting a brand new conversation when a draft isn't right, instead of continuing the one where Claude already has the context. A fresh conversation means re-explaining the audience, the constraints, and everything else from the first prompt, which is slower than pointing at what's wrong in the one you already have open.
This mistake is easy to make by accident. It's common to open a new chat out of habit, especially on a phone, without noticing that the conversation with the actual context is still sitting in your history. Before starting fresh, it's worth a quick check: is there already a conversation that has this context, and would continuing it be faster than restating everything?
There's one real exception. If the task has genuinely changed, not a correction to the same output but a different request entirely, a new conversation is the right call. The test is whether you'd still want Claude to remember the earlier context. If yes, stay in the same thread. If the earlier context would actually get in the way, start clean.
Continue reading
- Writing Prompts That Get You a Usable First Draft: the framework for a strong first prompt, so there's less to correct in the first place.
- Getting Claude to Ask You the Right Questions First: heading off the wrong guess before it happens, instead of fixing it after.
- Correcting a Fact Versus Correcting a Tone: a closer look at one specific, high-stakes case of refining.
- Uploading Files Claude Can Actually Use: applying the same specific-question habit to a task built around an uploaded file.