Why Claude Ignores Your Corrections + The Fix

“That’s wrong, try again” is the least useful correction you can give Claude. It produces an apology, a guess at what upset you, and a rewrite that breaks the paragraphs you liked. A correction that names the problem gets the real fix in one pass.

In this post, I will give you:

🎯 The three-part correction message that fixes anything in one pass.

🔀 The edit-and-branch move most users never touch.

📋 My copy-paste Fix It prompt that works on any bad output.

If you have not followed the Claude Genius podcast yet, go to your podcast app and do a quick search for “Claude Genius” and hit follow.

When you reply “that’s not right,” Claude learns exactly one thing: something displeased you. It does not learn what.

So it guesses. And the safest guess, from Claude’s side, is to change a lot. That is why a vague correction so often returns a draft where the problem is half fixed and the paragraphs you liked are gone.

Say Claude just drafted a blog post for you. The headline works, the intro sings, but the second section reads like it was written by a committee of lawyers.

Here is how to correct it:

#1 Name the exact problem. Quote the sentence. Point to the section. Describe what you see on the page, not the mood it put you in.

⛔ “The tone is off.”

✅ “The second section reads like a legal brief, full of words like ‘aforementioned’ and ‘pursuant.’ Match the plain voice of the intro.”

#2 Show one example of what you want. Anthropic’s docs treat examples as one of the most reliable steering tools, and a correction is where that pays off most. Paste a paragraph you like from somewhere else, or write two sentences in the voice you want.

⛔ “Make it punchier.”

✅ “Here’s the energy I want in that section: ‘Your inbox is not a to-do list. It just plays one on your phone.’ Rewrite it with that kind of snap.”

#3 Fence off what worked. Tell Claude what to keep before you tell it what to change. Almost everyone skips this step, and it is the one that stops the collateral damage.

⛔ “Fix the second section.”

✅ “Keep the headline, the intro, and the closing paragraph exactly as they are. Change only the second section.”Bonus: Make Claude grade its own fix.

Add one closing line to any correction:

✅ “After the revision, check it against my original request and list anything that still does not match.”

This comes straight from Anthropic’s prompting guidance, which recommends asking Claude to verify its work against your criteria before it finishes.

At the end of this post, I will give you my complete Fix It prompt template. It uses this same system to quickly fix anything that Claude came up short on.

The three-part correction fixes Claude's mistakes. These two moves are for when the mistake started with you, or the chat itself has gone stale.

Edit when the fault is in your request. If the mistake traces back to a wrong assumption, a missing constraint, or an instruction you never gave, the fix does not belong in a reply. Find your original message in the chat and click Edit.

Claude starts a fresh branch of the conversation from the corrected message. You get an answer built on the right foundation instead of a patch stacked on a misunderstanding.

You can always return to the original version, so you risk nothing by trying.

Start a new chat after two failed corrections. If you have corrected the same problem twice and it is still there, stop correcting. Claude is now rereading every failed attempt each time it answers.

Anthropic says a clean session with a better prompt almost always beats a long one full of accumulated corrections. So start fresh, and put everything you learned from the failed round into the first message.

You don’t have to write that first message yourself. Before you leave the old chat, send this prompt, then paste Claude's reply as the opening message of the new one chat:

Summarize this conversation as a handoff brief I will paste into a new chat, written as instructions to a fresh Claude that has seen none of it. Include: the goal I am trying to reach, stated in one sentence. The full current version of the output, even the parts that are wrong. A list of what I asked you to fix, what you tried, and why each attempt fell short. Every constraint, preference, and decision I gave along the way, including anything I told you to keep or avoid. End with the single most important instruction the fresh Claude needs to get it right on the first try. Give me only the brief, ready to paste, with nothing addressed to me.

Here’s the prompt template I use to quickly fix things that Claude came up short on.

The prompt below is for paid subscribers only.