How to Fix Bad AI Responses: A Diagnostic Guide

How to Diagnose a Bad AI Output

When an AI output disappoints you, the instinct is often to blame the AI. In reality, the vast majority of bad outputs are caused by something missing or unclear in the prompt. Before you can fix the problem, you need to identify which category of failure you are dealing with.

Fixing Generic, Vague Outputs

Generic outputs are the most common complaint about AI tools. The output is technically correct but could have been written about any version of the topic — it has no specificity, no personality, and no relevance to your particular situation.

The root cause is almost always missing context. The AI does not know who you are, who your audience is, what makes your situation unique, or what specific angle you want to take. Without this context, it defaults to the most common, middle-of-the-road version of the output.

[Add to your existing prompt]

Specific context that makes this unique:
- My situation: [what is specific about your case]
- My audience: [who exactly will read/use this, with specific details]
- My differentiator: [what makes this different from the generic version]
- My goal: [the specific outcome I want to achieve]

Avoid generic advice. Every point should be specific to this context.

Fixing the Wrong Tone

Tone mismatches are particularly frustrating because the content may be good but the delivery is completely wrong for the context. The AI defaults to a tone it considers appropriate for the task type — which may be formal when you need casual, or enthusiastic when you need measured.

Rewrite the previous output with the following tone adjustments:

Current tone problem: [describe what is wrong — too formal / too casual / too salesy / too academic / etc.]

Target tone: [describe exactly what you want]

Reference: [If possible, describe a publication, writer, or brand whose tone you want to match. E.g. "Write like a Wired magazine feature article" or "Write like a message from a trusted friend who happens to be an expert"]

Specific changes:
- Sentence length: [shorter / longer / varied]
- Vocabulary level: [simpler / more technical / conversational]
- Point of view: [first person / second person / third person]
- Energy level: [calm and measured / energetic and direct / warm and encouraging]

Fixing Factual Errors and Hallucinations

AI hallucination — where the model confidently states incorrect information — is one of the most serious failure modes. It is particularly dangerous for tasks involving statistics, dates, names, technical specifications, or legal information.

The best defence against hallucination is a combination of explicit instructions and verification prompts. Instructing the AI to flag uncertainty, cite sources, and distinguish between known facts and inferences significantly reduces the risk of confident misinformation.

[Add to any research or factual prompt]

Important instructions for accuracy:
1. Only include information you are confident is accurate
2. For any statistic or specific claim, note the approximate source or context (e.g. "according to a 2023 McKinsey report" or "widely cited figure, recommend verification")
3. If you are uncertain about any fact, say so explicitly rather than guessing
4. Distinguish clearly between established facts and your analysis or inference
5. Note your knowledge cutoff date where relevant to the topic
6. Flag any areas where you recommend I verify the information independently

Fixing Format and Structure Problems

Format problems occur when the AI produces the right content in the wrong structure. This might mean bullet points when you wanted prose, a wall of text when you wanted sections, or a formal report structure when you wanted a conversational email.

Reformat the previous output with the following structure:

[Choose the format you need]

Option A — Document structure:
- Section 1: [heading] — [length]
- Section 2: [heading] — [length]
- Section 3: [heading] — [length]

Option B — Prose vs lists:
- Use flowing paragraphs for [sections]
- Use bullet points only for [sections]
- No more than [number] bullet points per section

Option C — Length adjustment:
- Target length: [word count]
- Cut: [what to remove or condense]
- Expand: [what needs more detail]

Keep all the content — only change the structure.

Using the Refinement Loop

Professional AI users rarely accept the first output. They use a refinement loop — a series of targeted follow-up prompts that progressively improve the output until it meets their standard. This is not a sign that the AI failed; it is how professional-quality outputs are produced.

Review the output you just produced and answer these questions honestly:

1. What are the three weakest or least convincing parts?
2. Where is the reasoning thin, unsupported, or generic?
3. What important information is missing that a knowledgeable reader would expect?
4. What would a harsh editor cut or rewrite?

Then produce an improved version that addresses all four points.

⚡ RESEARCH DIRECTOR — OUTPUT QUALITY REVIEW MODE

REVIEW BRIEF:
- Output type: [article / report / analysis / email / code / etc.]
- Quality standard: [publication / client-facing / internal / personal use]
- Primary audience: [who will read or use this]

REVIEW CRITERIA:
Rate each dimension 1–10 and explain:
1. Accuracy: Are all facts, figures, and claims correct and verifiable?
2. Completeness: Is anything important missing?
3. Clarity: Is the writing clear and easy to follow?
4. Specificity: Are there vague generalisations that should be made concrete?
5. Tone: Does the voice match the audience and purpose?
6. Structure: Is the organisation logical and easy to navigate?

IMPROVEMENT INSTRUCTIONS:
For each dimension scoring below 8, provide specific rewrite instructions.
Then produce the improved version.

When to Start Fresh vs When to Refine

Sometimes the most efficient path is to abandon a conversation and start fresh with a completely rewritten prompt, rather than trying to refine an output that went badly wrong from the start. Knowing when to do this saves significant time.

Start fresh when: the output is fundamentally wrong in its approach or framing, the AI has made an incorrect assumption that is now embedded in the conversation context, or you have spent more than three refinement rounds without meaningful improvement. In these cases, the conversation context itself may be working against you.

Continue refining when: the output has the right structure but needs content improvements, the tone is close but not quite right, or specific sections need to be expanded or condensed. Targeted refinement prompts are faster than starting from scratch when the foundation is sound.