Quick answer
When you know an AI result is wrong but can’t say why, stop trying to write the perfect instruction and act as the judge instead. React to specific parts (“keep this, lose that”), compare it with examples you like, ask the AI to show you a few clearly different options, and let it ask you short either/or questions. Then have it write down what you want and don’t want, and check that list before the next attempt. Keep every rejected version; together they teach the model your taste.
📘 Main article: This guide builds on our article Give AI Your Failed Attempts: How Fable 5.2 Got It Right First Time, where Nate B Jones admits his feedback was often just “it doesn’t work, fix it”, yet Claude Fable 5.2 still worked out what he meant.
Key takeaways
- Recognising is easier than describing. Use that: react rather than specify.
- Point at parts, not the whole thing.
- Compare options side by side to find out what you actually prefer.
- Make the AI write the rules it has learned from you, then check them.
In this guide
- Why is it hard to explain what is wrong?
- Why act as the judge, not the designer?
- How do you give useful AI feedback? A 5-step method
- What can you say instead of “it’s not right”?
- A prompt that makes the AI do the explaining
- When should you stop and start again?
- Frequently asked questions
- The bottom line
- Related posts
“That’s not it.” Most of us have typed something like this into an AI chat, then watched it produce another version that is wrong in a new way. It is frustrating, but it is also normal. Lots of real work, such as design, writing, page layouts and branding, is “I’ll know it when I see it”.
In the main article, Nate B Jones describes exactly this. He is picky about design but not a designer. He could spot what was wrong far more easily than he could explain what he wanted. This guide turns that situation into a repeatable process.
Why is it hard to explain what is wrong?
Your sense of “right” is built from experience, taste and goals you have never written down. When something is off, you feel it before you can name it. Vague feedback is not a sign you are bad at prompting. It just means the missing piece is still in your head.
The trick is to give the AI ways to get at that missing piece without needing a perfect brief from you.
Why act as the judge, not the designer?
Nate says he deliberately engaged with the model “as a customer” on design, because that is where he could drive quality best. He knows product and technical work well, but not design, so he judged rather than directed.
This is good news for small business owners. You do not need to know design terms or copywriting rules to get good results. You need to be clear about what works for you and your customers, and honest when something does not.
💡 Tip: Before you start, write one sentence on who the work is for and what it must help them do. Judge every version against that sentence, not against your mood.
How do you give useful AI feedback? A 5-step method
- Point at parts. Instead of “it’s wrong”, go through the result and mark each part keep, change or remove. Even “the top half works, the bottom half doesn’t” is useful.
- Describe the feeling, then the effect. “It feels cluttered, so I can’t find the date quickly” gives the AI a problem to solve.
- Show an example. Share a screenshot, page or piece of writing you like, and say which bit you like. Examples often say more than adjectives.
- Ask for contrasting options. Get three clearly different versions and pick the closest. Choosing is much easier than describing.
- Make it write the rules. Ask the AI to sum up what it has learned about your preferences as a short “want / don’t want” list. Correct it, then keep it for next time.
⚠️ Warning: Do not delete rejected versions. Keep them in a folder with a line on what was wrong. As Nate found, a trail of failed attempts and blunt feedback can later help a stronger model work out what you were after.
What can you say instead of “it’s not right”?
| Instead of… | Try… |
|---|---|
| “It’s not right.” | “The layout is fine, but the headline doesn’t stand out. Make it the first thing I see.” |
| “It looks cheap.” | “It feels busy. Use fewer colours and more space, like this example.” |
| “The tone is off.” | “It sounds like a salesperson. Make it sound like a helpful friend explaining something.” |
| “Make it better.” | “Keep sections 1 and 3. Rewrite 2; it repeats what 1 already says.” |
| “I hate it.” | “This is the opposite of what I want. Go in the other direction on colour and length.” |
| “Fix it.” | “Show me three versions: one simpler, one bolder, one more formal.” |
You will not always find the right words, and that is fine. Even a rough “warmer” or “calmer” narrows things down when paired with a specific part.
A prompt that makes the AI do the explaining
When you are truly stuck, hand the explaining over to the AI with this prompt:
I don't like this version, but I can't fully explain why.
Help me work it out:
1. Show me three quite different alternatives (A, B, C), each changing one big thing.
2. For each, say in one line what you changed and why.
3. Then ask me up to three yes/no or either/or questions about what I prefer.
After I answer, write a short list of "what I want" and "what I don't want"
and check it with me before making the next version.
The either/or questions are the key. Answering “bigger date or bigger weather?” is far easier than writing a design brief, and a few answers often reveal what you really care about.
See the full story: read how Claude Fable 5.2 read a month of failed attempts and produced a design Nate liked straight away.
When should you stop and start again?
Sometimes another round in the same chat will not help. Start again when:
- The AI keeps bringing back things you rejected.
- Versions are getting worse, not closer.
- The chat is so long it has started forgetting earlier decisions.
When that happens, do not throw the work away. Gather the attempts and your feedback, then either write a handoff file for a fresh chat or give the whole trail to a stronger model, as described in the main article. For repeatable jobs, a written AI feedback loop with clear checks stops the same problems coming back.
Frequently asked questions
How do I give AI feedback if I can’t explain what’s wrong?
React to specific parts, describe how it feels and what effect that has, show examples you like, and ask the AI for a few contrasting options and short either/or questions.
Should I start a new chat when the AI keeps getting it wrong?
If it keeps repeating rejected ideas or the chat is very long, yes. Carry over a short summary of decisions and dislikes so you do not lose what you have learned.
Is it worth keeping AI drafts I rejected?
Yes. Rejected versions plus your notes on what was wrong help a model understand your preferences in a later attempt.
Do I need design or writing skills to judge AI work?
No. You need to know your customers and what the work must achieve. Judging against that is a skill most business owners already have.
What is a “want / don’t want” list?
A short list the AI writes summarising your preferences from your feedback. You correct it, then reuse it on future tasks so you do not repeat yourself.
The bottom line
You do not need perfect words to get good AI results. Act as the judge: point at parts, show examples, choose between options and let the AI ask the questions. Write down what it learns about your taste, keep your rejected versions, and each attempt gets you closer.
Your next step: next time you think “that’s not it”, paste the prompt above instead of starting again. Then read the main article: Give AI Your Failed Attempts: How Fable 5.2 Got It Right First Time.
Related posts
- Give AI Your Failed Attempts: How Fable 5.2 Got It Right First Time
- How to Build an AI Agent Feedback Loop: Brief, Check, Score, Repeat
- How to Lock a Design System Before Your AI Agent Writes Code
