ChatGPT Not Giving You What You Want? 12 Fixes

When ChatGPT misses, the instinct is to retype everything from scratch or to argue with it. Neither is efficient. Almost every disappointing answer has a recognizable shape, and each shape has a short repair you can paste into the same chat.

This article lists twelve symptoms with the exact prompt for each. Replace anything in [brackets] with your own details.

Diagnose the symptom before you retype anything

Use three steps every time.

1. Name the symptom in a few words: "too generic," "ignores my length limit."
2. Find the likely cause. Usually something was missing from your original prompt: a reader, a fact, a rule, an example. If you want a refresher on those parts, see how to write better ChatGPT prompts.
3. Repair one thing at a time. If you change four things at once, you will not know which one helped.

Also reread your own prompt first. Sometimes the answer is a faithful reply to what you actually wrote.

Fixes 1 to 5: generic, off-tone, ignored rules and wrong length

1. The answer is too generic. It had only generic input, so it gave generic output. Add a specific reader and real facts.

That is too generic. Here is what you were missing: the reader is [specific person and what they care about]. The key facts are [facts]. The one thing this must achieve is [outcome]. Rewrite it so nothing in it could appear in a version written for a different reader.

2. The tone is wrong. Words like "friendlier" are vague. Show a sample sentence instead.

The tone is off. Aim for formality 2 out of 5 and warmth 4 out of 5. It should sound like this sentence: "[a sentence in the voice you want]". Change only the wording, not the content, and avoid [words or habits that bother you].

3. It ignored a rule. Rules buried in a long prompt get missed. Isolate the one that matters.

You missed one rule: [state the single rule]. It is the most important requirement. Revise the answer so it follows that rule exactly. Before replying, check your draft against the rule and tell me in one line whether it passes.

4. It is too long. Give numbers and tell it what to cut first.

Cut this to between [X] and [Y] words. Keep [the two or three things that must survive]. Remove repetition and background first. Do not add anything new. Tell me the final word count.

5. It is too short or too thin. Say what depth looks like.

This is too thin. For each point, add one concrete example and one sentence on what to do about it. Keep the structure. Aim for roughly [range] words.

Note that assistants are often imprecise about exact word counts. Treat a requested count as a target, and check it yourself if it matters.

Fixes 6 to 8: made-up facts, repetition and long-chat drift

6. It made things up. Assistants can produce confident, wrong details, especially when you did not supply the source. Force it to separate what you gave from what it guessed.

Some details in your answer do not appear in what I provided. Go through your answer and label each factual claim SUPPLIED (from my text), GENERAL (widely known, but I should still verify), or UNSURE. Then rewrite using only SUPPLIED claims and list what you removed.

This does not make the output trustworthy by itself. You still need to check names, numbers and quotes yourself. The verification checklist shows how.

7. It repeats itself. Repetition is usually a structure problem, not a wording problem.

The sections overlap. List each distinct point once in a numbered list, assigned to the single section where it fits best. Then rewrite so every point appears exactly once.

8. It drifts in a long chat. After many turns, early instructions get diluted and contradictions pile up. Restate the brief.

We have wandered, so reset. This brief replaces all earlier instructions. Goal: [goal]. Reader: [reader]. Rules: [rules]. Format: [format]. Acknowledge in one line, then redo the last task under this brief only.

If a reset does not hold, open a new chat and paste a cleaned-up version of your best prompt. A fresh chat is often faster than rescuing a cluttered one.

Fixes 9 to 12: refusals, misreadings, flattery and format

9. It refused. First decide whether the refusal is right. If you asked for something harmful or unsafe, the refusal is doing its job and you should not try to get around it. If your request was legitimate and simply misread, clarify honestly.

I think my request was misunderstood. My actual purpose is [honest purpose and context]. I am not asking for [what it may have assumed]. What I need is [the safe version of the request]. If you still cannot help, tell me which part you cannot do so I can handle it another way.

10. It answered a different question. Ask it to say what it understood before it answers.

Before answering, restate my request in two sentences and list any assumptions you are making. Wait for my confirmation, then continue.

11. It agrees with everything. If you ask "is my plan good?" you tend to hear yes. Ask for the opposite job.

Act as a skeptical reviewer. List the three strongest reasons this might not work, the assumption I am most likely wrong about, and one thing I should check before committing. Do not soften the criticism and do not offer praise.

For more prompts in this spirit, see thinking-partner prompts like devil's advocate and pre-mortem.

12. The format is wrong. Describing a format in words is easy to misread. Give a skeleton.

Reformat your answer to match this skeleton exactly, with no extra commentary:
Title: [one line]
Summary: [two sentences]
Steps: [numbered list, each under 15 words]
Risks: [two bullets]

The order that saves the most time

When an answer fails, work down this list. It is ordered roughly from cheapest to most disruptive, and it is a rule of thumb rather than a law.

1. Reread your prompt and make sure the answer is not simply correct for what you wrote.
2. Add the missing fact, reader or example.
3. Restate the broken rule by itself.
4. Ask for a self-check against your own criteria.
5. Start a fresh chat with a cleaned-up prompt.

Keep the repair prompts that work. A short list of your own, saved in a notes file, will cut your frustration more than any single trick.

Where to find more

These twelve fixes are a small slice of a bigger toolkit. The Prompt Bible has a full chapter on diagnosing wrong answers, plus format control, iteration habits and reusable patterns. If you want more advanced techniques such as thinking partners, verification routines and learning prompts, Secret Prompts: 75 Power Techniques Most People Never Use collects them in one place.

Frequently asked questions

Why does ChatGPT ignore my instructions?

Rules get lost when a prompt is long, when rules conflict, or when a chat has run on for many turns. Restate the one rule that matters, rank your rules, and ask for a self-check before it replies.

Should I start a new chat or keep correcting the old one?

Correct in place for small misses. If the chat is long or the answers keep drifting, start fresh with a cleaned-up prompt that includes the current brief.

Can I stop ChatGPT from making things up completely?

No. You can reduce it by pasting source text, allowing "I do not know," and asking it to separate supplied facts from guesses, but you still need to verify anything important.

Why do I get different answers to the same prompt?

Assistants generate text with some randomness, so wording and content can vary between runs. If a result matters, ask for several versions and choose the best.