Claude rarely gets a non-trivial change perfectly right on the first try. That’s not a defect — that’s how this works. The mistake isn’t the first attempt; it’s how people respond to it.
Maybe you already rejected the diff because it wasn’t right. Maybe you accepted it and only then noticed. Either way, the next move is the same: don’t start over — say what’s still wrong.
When Claude is almost right, the worst move is starting over. Tell it what’s still wrong.
The whole point of working with Claude inside a single chat is that the context builds up. Claude remembers what you asked for, what it just proposed, what files it looked at, what tradeoffs it considered. Restart and all of that is gone. You’re starting from a cold deck for the second time — except now you’re frustrated.
The bad follow-up
The most common bad follow-up is no follow-up at all — just rage-deleting the chat and starting over with a slightly different prompt.
The second most common is so vague it doesn’t help:
- “no, try again”
- “that’s not what I meant”
- “hmm, can you redo it”
Claude has no idea what part was wrong. So it’ll change something — probably the wrong something — and you’ve burned a turn.
The good follow-up
A good follow-up does two things:
- Says what’s still wrong. Specifically. Not “it’s broken” — “the button is now blue, but I asked you to keep the green from before.”
- Says what you want instead. “Change the color back to green; leave the padding and font as they are.”
That’s it. Two sentences. Notice what’s not in there: a re-explanation of the whole task. Claude already knows the task. Just tell it the delta.
A worked example
You asked Claude to fix the contact form in contact.html so the error message shows up below the message box instead of above it.
The diff Claude proposes moves the error message and changes its color to red. You read it, see the surprise color change, and reply — whether you’ve already accepted or you’re about to:
- Bad reply: “no the colors are wrong, redo it”
- Better reply: “The position is right — keep that. But don’t change the color; it should stay the gray it was before.”
Claude now knows: keep your work, change one specific thing, leave the rest alone.
When to actually restart
Iteration is the default. Restart only when:
- Three or four rounds in and you’re going backward — the diff is bigger and uglier each time.
- The original prompt was based on a wrong assumption and the whole approach needs rethinking.
- You can’t remember what the goal was anymore (this happens).
The last one is the most honest signal. We’ll come back to it in Lesson 4.8.
What’s next
Good follow-ups are how you fix vague prompts after the fact. The next lesson is how to skip the vagueness in the first place.