What is it?
Prompts fail in predictable ways. Because the model can only act on the text it receives, most "bad AI answers" are really under-specified requests. This page lists the common failures with a fix for each, so you can debug a prompt systematically.
Why does it matter?
Debugging by guesswork wastes time and leads to superstition ("it works if I say please in capitals"). Naming the failure mode lets you make a targeted change and test it. In applications, the same mistakes cause inconsistent behaviour that is expensive to find later.
How does it work?
Symptom → likely cause → fix:
| Symptom | Likely cause | Fix |
|---|---|---|
| Generic, bland answer | Goal and audience not stated | State who it is for and what decision or action it supports |
| Answer ignores your data | Data missing, buried, or too long | Include the relevant text, delimit it, trim the rest (What is a Context Window?) |
| Wrong format | Format only implied | Specify the exact format; add a short example |
| Invented facts | No source and no permission to say "unknown" | Supply sources and allow "not found" (How to Reduce AI Hallucinations) |
| Instruction skipped | Too many rules, or rules conflict | Prioritise, remove duplicates, split into steps |
| Inconsistent across runs | Ambiguity or high randomness | Tighten wording, add examples, lower temperature if available |
| Works on one test, fails on others | Tested once | Build a small set of varied cases (How to Evaluate AI Systems and Outputs) |
Example
Before: "Write something about our new feature."
Problems: no audience, no length, no format, no facts about the feature.
After:
Write a 120-word announcement for existing customers about the new CSV export.
Facts: available on all paid plans; exports up to 50,000 rows; found under Reports > Export.
Tone: friendly, no hype. End with one sentence linking to the help article.
Do not mention features not listed above.Every sentence removes a guess the model would otherwise make.
When should I use it?
Run through this list whenever a prompt underperforms, before you conclude the model is "not good enough". If a well-formed prompt still fails, the problem may be missing knowledge (What is RAG (Retrieval-Augmented Generation)?), a task that needs a tool (Function Calling (Tool Calling) Explained), or a task better solved by ordinary code.
Common mistakes
Beyond the table above:
- Only saying what not to do. "Don't be formal" is weaker than "write in a casual, conversational tone".
- Over-engineering magic phrases. Clear instructions beat tricks.
- Changing many things at once, so you cannot tell which change helped.
- Pasting confidential data into prompts without considering where it goes (AI Privacy and Security: What to Know Before You Paste Data).
Try it with GoVibeTool
Several of these mistakes — missing context, no stated output format, no tone — disappear when each part has its own box. The AI Prompt Builder is one way to check you have filled them in before sending.
Runs in your browser; your draft is not sent to a server.