Guide

Why ChatGPT keeps getting it wrong, and how to fix it

Published 3 October 2026

A chat window with a greyed-out generic reply next to a notebook listing what the prompt was missing

Short answerMost bad ChatGPT answers come from a prompt that's missing context, a format you never stated, or a long chat where earlier instructions got buried. Fix the specific cause instead of just asking again, and the next answer is usually right.

ChatGPT answered, and it's not what you needed. Before you try the same prompt again, or give up and write it yourself, look at what the answer tells you about the prompt. Most problems fall into a short list, and each one has a direct fix. This is the text-prompt version of our guide to why Midjourney prompts don't work.

1. The answer is generic

Symptom: it could have been written for anyone, about anything.

Fix: a generic prompt gets a generic answer. Give ChatGPT a role, an audience and real material, instead of just a topic. The role prompt template does this directly:

prompt — 5 blanks

Act as a [ROLE, e.g. senior HR manager] with [NUMBER] years of experience working with [WHO, e.g. small businesses under 50 staff].

My situation: [YOUR SITUATION, SPECIFIC DETAILS]
What I need: [THE DECISION OR DOCUMENT YOU NEED]

Ask me anything you need to know before you answer, rather than guessing.

The last line matters as much as the role. Letting the model ask first stops it from filling gaps with generic material.

2. It makes up facts

Symptom: a number, a quote or a feature that turns out not to exist.

Fix: ChatGPT will fill a gap in what it knows with something plausible unless you tell it not to. Say so directly: "If you're not certain, say so instead of guessing" or "search for this before you answer, and give me the source". Never let a figure you plan to use land in a document unchecked. Treat any number, date or name it gives you as a draft to verify, not a fact.

3. It sounds like an AI wrote it

Symptom: long sentences, a lot of bold text, phrases like "in today's fast-paced world".

Fix: ask for the opposite directly, and give it a sample of how you actually write. The humanise AI text prompt is built for exactly this:

prompt — 2 blanks

Rewrite the text below so it sounds like a real person wrote it.

- Vary sentence length. Use some short, punchy sentences.
- Use simple, specific words. Remove filler and hedging.
- No clichés, no "unlock" or "game-changer" style phrases.

Here's a paragraph I wrote, to match the voice: [PASTE A SAMPLE OF YOUR OWN WRITING]

Text to rewrite: [PASTE TEXT]

Giving it your own writing as a sample works far better than describing your tone in adjectives. "Casual but professional" means something different to everyone; a real paragraph doesn't.

4. It forgot something from earlier in the chat

Symptom: an instruction you gave ten messages ago has quietly stopped applying.

Fix: in a long conversation, earlier instructions can get crowded out by everything said since. Don't assume it remembers. Repeat the constraint that matters in the same message as your next request: "keep this under 200 words, like the last few replies" rather than relying on it to recall your original brief. For anything you'll want to reuse, turn the prompt into a saved template once it works, rather than re-explaining your requirements every time. The reusable template prompt does this for you:

prompt — 2 blanks

Here is a prompt that worked well for me:

[PASTE PROMPT]

Turn it into a reusable template with [PLACEHOLDERS] for the parts that change, and a one-line note on what each placeholder should contain.

5. The format or length is wrong

Symptom: a wall of text when you wanted bullets, or three sentences when you wanted a full breakdown.

Fix: state the exact shape you want, not just the topic. "Summarise this" leaves the format up to the model. The executive summary prompt states it precisely:

prompt — 1 blank

Write an executive summary of the document below for a busy senior reader.

Document: [PASTE TEXT]

Format: one-line bottom line first, then 5 bullets of the most important points, then one sentence on what to do next.

If a reply still comes back the wrong shape, the fix is a follow-up, not a new prompt: "same content, but as a numbered list of 5 items" gets there faster than starting again.

6. It refuses, or gives you a watered-down answer

Symptom: a flat refusal, or an answer so hedged it's not useful.

Fix: check whether the request reads as sensitive out of context. Add the legitimate reason plainly: who it's for and why, rather than just the ask itself. If the topic is genuinely sensitive, a different model or a narrower, more specific question usually gets further than repeating the same broad one.

7. It did more than you asked

Symptom: you asked for one paragraph and got a rewritten document with three unrequested sections.

Fix: say the boundary out loud: "Do only what I asked. If you notice something else worth changing, list it at the end as a suggestion, don't do it." This one line stops most scope creep without you having to catch it after the fact.

A routine for any bad answer

Name what's actually wrong in one sentence: "too generic", "wrong length", "sounds robotic".
Match it to the cause above.
Fix that one thing with a follow-up message, not a brand new prompt.
If it's still wrong after two follow-ups, check whether the original prompt gave it enough real material to work with. Thin input is the root cause more often than the model.

Where to go from here

For editing a prompt once you've found the fix, see how to edit an AI prompt and how to use a prompt template. Browse more text prompts in chat and writing prompts, or use the prompt builder to start from a template with the placeholders already in place.