When AI gets something wrong, first name the kind of error, then fix the prompt or the source material and regenerate, rather than arguing in the same chat. A wrong fact, an ignored instruction and a bland tone are three different problems, and each has a different fix.
This post gives you a diagnosis table, a recovery loop with an example, what the research says about asking AI to correct itself, and a checklist so the same mistake doesn't come back.
What kind of wrong is it?
Match the symptom to the failure type before you touch anything. Most people reply "no, try again", which treats every failure as the same one.
| Symptom | Likely cause | First fix |
|---|---|---|
| A claim, date, quote or statistic is wrong or invented | The model generated plausible text without a source | Supply the facts; allow "I don't know"; verify outside the chat |
| It skipped or bent an instruction | Instruction buried, vague, or conflicting | Move it to the end, make it specific, remove conflicts |
| Right content, wrong tone or reading level | No audience or voice sample | Name the reader and add a sample of the voice you want |
| Wrong shape (missing sections, wrong length, wrong format) | Format described loosely | Show an example of the exact format |
| Generic, could-be-anyone answer | Prompt had no specifics | Add your notes, constraints and one real example |
| It rewrote more than you asked | No edit boundary | State the edit level and protect list |
Those causes are my working taxonomy, drawn from the documentation and studies cited below, not an official classification. The point is the habit: diagnose, then fix.
Should I argue with the AI or edit the prompt?
Edit the prompt and regenerate. Arguing in the same thread tends to add noise. A long back-and-forth fills the context with your corrections and the model's apologies, and research suggests extra text can hurt: in a 2023 study (Shi et al.), adding irrelevant context to math problems dramatically reduced accuracy. Our post on prompt length covers a second study where accuracy dropped from 0.92 to 0.68 as filler grew to 3,000 tokens.
There's another reason to avoid pushing back on facts. Sharma et al. (2023) studied five AI assistants and found that sycophancy, a tendency to match what the user seems to believe, shows up consistently, and that both humans and preference models sometimes prefer a convincing agreeable answer over a correct one. If you say "I think the answer is X" the model may follow you there whether or not X is right.
A cleaner loop looks like this:
- Copy the original prompt into a notes file.
- Write the failure in one line ("it invented a 2019 statistic").
- Add the fix to the prompt itself, in a place where it can't be missed.
- Start a fresh chat, run the revised prompt, and compare.
What does a recovery look like in practice?
It looks like a smaller, more specific second prompt. Say you asked for a chapter on saving for taxes and the draft included "According to a 2019 survey, 62% of freelancers under-save for taxes." You don't have that survey. Here's a weak response and a better one:
Weak: That statistic isn't real. Try again and be more accurate.
Better: Rewrite the chapter using only the notes below. Do not include any statistic, survey, quote or date that is not in my notes. If a point would benefit from data, write [DATA NEEDED: what kind] and continue.
NOTES:
- Many freelancers set aside a fixed percentage of each invoice for taxes.
- Pay estimated taxes quarterly if your country requires it; check with a tax professional.
The "better" version changes the rules (use only my notes), gives the model a safe exit instead of forcing it to invent (the placeholder), and supplies the facts. Anthropic's documentation on reducing hallucinations recommends the same ingredients: let the model say it doesn't know, restrict it to the provided material, and verify claims against sources. It also says plainly that these techniques significantly reduce hallucinations but don't eliminate them. Our full guide to stopping AI from making up facts expands on this.
Can the AI fix its own mistakes?
It can fix checkable ones; for reasoning and facts you need outside feedback. In "Large Language Models Cannot Self-Correct Reasoning Yet" (Huang et al., 2023), the authors found that without external feedback, models struggle to correct their reasoning and sometimes perform worse after trying. That doesn't mean review prompts are useless. Other research shows structured verification can help when it's designed well: Chain-of-Verification (Dhuliawala et al., 2023) has the model draft an answer, plan verification questions, answer those questions independently, and then write a final response, and it reduced hallucinations in their tests.
You can borrow the idea in plain language:
Step 1: Draft the section from my notes.
Step 2: List every factual claim in your draft as a numbered list.
Step 3: For each claim, mark it SUPPORTED (quote the line from my notes) or UNSUPPORTED.
Step 4: Rewrite the section removing or flagging every UNSUPPORTED claim.
That works because "is this claim in my notes?" is a question with an answer on the page. "Is this true?" is not. You are still the final check on anything a reader might rely on.
Separate steps matter too. Anthropic and Google both recommend chaining prompts so you can inspect each stage; in a 20-person study of that approach (Wu et al., 2021), chaining improved quality, transparency and control. We walk through the beginner version in prompt chaining for planning, drafting and editing. The same chain is built into Interactive eBook Builder as three separate agents (Spec Builder, Drafter, Polisher), so a mistake can be traced to the stage that made it.
How do I stop the same mistake from returning?
Turn each recurring error into a permanent line in your setup. If the AI keeps writing 900-word chapters when you asked for 600, that is a standing rule. If it keeps adding "In conclusion", that is a standing rule. Keep a short log:
| Error seen | Rule added |
|---|---|
| Invented statistics | Use only supplied notes; use [DATA NEEDED] placeholders |
| Chapters ran long | "Between 550 and 650 words" |
| Recap paragraph at end | "End on the last step; no summary" |
| Stiff tone | Added a 200-word voice sample |
If you use a structured workflow, this log can live in the tool. The Packs Database in Interactive eBook Builder holds QA/Gate entries alongside things like Constraint Packs, so a check you invented after one bad chapter can apply to the next. And when a workspace itself gets tangled, the Seed Prompt can be run again to repair it, with the final QA pass in Agent C Polisher catching what slipped through. You can also keep the same habit with nothing but a notes file.
If your answers are wrong mostly because they're vague, the fix is different, and how to test and improve a prompt when AI gives generic answers covers it. For instructions that are simply being skipped, see why AI ignores your instructions.
Before you reply "try again"
- Did you write down what kind of wrong it was?
- Can the fix go into the original prompt instead of a reply?
- Did you give a safe way out ("say you're unsure", a placeholder)?
- Will you verify any fact against a source you picked?
- Did you add a line to your rules so it doesn't recur?
The takeaway: treat a bad answer as a diagnosis problem. Fix the cause in the prompt, verify facts outside the chat, and the same mistake gets rarer.