The Recovery Guide — Fixing Bad AI Outputs
A practical guide to recovering from every type of AI failure — hallucination, wrong format, off-topic responses, and more.
The Recovery Guide 🎯
The correction is always faster than the restart.
Recovery by Failure Type
Hallucinated Facts
The AI confidently states something that is not true. This is the most common and most dangerous failure.
Detection: If a claim feels too specific, too convenient, or too perfectly aligned with your question — verify it. AI hallucination often looks plausible.
Recovery prompt:
"The statistic you cited about [X] appears to be incorrect. Provide this section again using only verifiable facts. If you are not confident in a specific number, say so rather than inventing one."
Prevention: Add "cite your sources" or "only include facts you are confident about" to your original prompt.
Wrong Format or Structure
The AI produced content but in the wrong format — a paragraph when you wanted bullet points, a list when you wanted a table, prose when you needed JSON.
Recovery prompt:
"Good content, wrong format. Reformat this as [specific format]. Keep all the information, change only the structure."
Prevention: Always specify format explicitly: "Respond as a markdown table with columns: X, Y, Z"
Off-Topic Response
The AI answered a different question than the one you asked, or focused on the wrong aspect.
Recovery prompt:
"That answers a different question. I am asking specifically about [X], not [Y]. Focus only on [X] and ignore [Y] entirely."
Prevention: End prompts with "Focus specifically on [X]" and "Do not discuss [Y]."
Wrong Tone or Register
The output is too formal, too casual, too verbose, or too academic for your needs.
Recovery prompt:
"Same content, different tone. Rewrite this as if explaining to [target audience]. Less [current problem], more [desired quality]."
Prevention: Include "Tone: [specific]" and "Audience: [specific]" in every prompt.
Fundamentally Wrong Approach
The AI took the wrong strategy entirely — wrote a comparison when you needed a how-to, built a class when you needed a function.
Recovery prompt:
"Wrong approach entirely. Do not revise — pivot. I need [describe correct approach] instead of [what it did]. Start fresh on the approach but keep the context we have established."
The Correction Escalation Ladder
When a simple correction does not work, escalate:
| Level | Technique | When to Use |
|---|---|---|
| 1. Nudge | "Adjust X to Y" | Minor issues |
| 2. Redirect | "Focus on A, not B" | Off-topic drift |
| 3. Correct | "X is wrong. The correct answer is Y" | Factual errors |
| 4. Reframe | "Let me rephrase: I need..." | Misunderstood intent |
| 5. Pivot | "Scrap that approach. Instead..." | Wrong strategy |
| 6. Reset context | "Forget the above. New task:" | Conversation poisoned |
| 7. New conversation | Close and start fresh | Unrecoverable mess |
Most failures resolve at Level 1-3. If you reach Level 6, consider whether the original prompt was the problem.
The 60/40 Rule
When evaluating a bad output, ask: is at least 60% of this usable? If yes, recover. If no, pivot or restart.
60%+ usable — correct the rest:
"The structure and first three sections are good. Rewrite sections 4 and 5 — they are [specific problem]."
Under 60% usable — pivot:
"This is not in the right direction. Let me restate what I need: [clear, specific restatement]."
Recovery Prompts That Always Work
These five prompts handle 90% of recovery situations:
- "What you wrote is partially correct. Keep [X, Y] but change [Z] because [reason]."
- "You misunderstood. I need [restate clearly]. Do not include [what went wrong]."
- "Same information, different format: [specify exact format]."
- "This is too [vague/detailed/formal/casual]. Rewrite at [length] in [tone]."
- "Review what you wrote. What are the weakest parts? Fix them."