The 7 Prompting Mistakes that Sabotage Your ChatGPT Results

Table of Contents

1. Mistake 1: the Prompt with No Context
2. Mistake 2: Asking for Everything at Once
3. Mistake 3: Not Specifying the Expected Format
4. Mistake 4: Forgetting to Assign a Role
5. Mistake 5: Accepting the First Answer
6. Mistake 6: Sharing Confidential Data
7. Mistake 7: Never Checking the Facts
8. Conclusion
9. FAQ

In French-speaking Switzerland, nearly three out of four workers already use AI on the job. Yet most of them never learned how to phrase a request properly. The result: generic answers, endless rephrasing and that hasty conclusion that “AI just doesn’t get it”. In reality, the quality of an answer depends first on the quality of the question. That is the whole point of prompt engineering, the discipline of structuring your instructions to get reliable results. The good news is that the mistakes are always the same. We have identified seven of them, observed across hundreds of training sessions. Fix them and your results will change today, whatever the tool: ChatGPT, Claude, Gemini or Copilot. Let us walk through them, each with a before-and-after example.

Mistake 1: the Prompt with No Context

“Write me an email for a client.” That is the most common prompt in the world, and also the least effective. Indeed, the model knows nothing about your situation: which client, which goal, which tone, which business relationship. Lacking information, it therefore produces a one-size-fits-all answer.

The fix is simple: give context before the instruction. Who you are, who you are writing to, what happened before, what you want to achieve. Compare for yourself. Before: “Write a follow-up email.” After: “I am a sales manager at a Geneva-based fiduciary. A prospect requested a quote for an accounting mandate ten days ago and never replied. Write a short, friendly follow-up that proposes a fifteen-minute call this week.” The second version produces an email you can send immediately. In other words, every minute invested in context saves you ten in corrections.

Mistake 2: Asking for Everything at Once

The second counterproductive habit: the mega-prompt that requests an analysis, an outline, a full draft and a translation in the same sentence. The model then tries to handle everything, and handles everything superficially. Moreover, you can no longer adjust course along the way, since everything arrives in one block.

Choose a step-by-step approach instead. First the structure: “Suggest a five-part outline for this report.” Then the adjustment: “Expand part 2 with a focus on risks.” Finally the polish: “Rewrite this paragraph in a more direct tone.” This cascade method keeps control in your hands. Besides, it matches the way models reason best: one clear instruction at a time. Complex tasks are handled as a conversation, not a monologue.

Mistake 3: Not Specifying the Expected Format

You wanted a comparison table, you got three paragraphs. You hoped for an action list, you received an essay. This mismatch is not inevitable: the model does not guess the format, you have to state it.

Get into the habit of ending your prompts with an explicit formatting instruction. A few phrases that change everything:

• “Present the result as a table with three columns: criterion, option A, option B”;
• “Answer in five bullet points maximum, one sentence per bullet”;
• “Write 150 words at most, professional tone”;
• “Give only the list, no introduction and no conclusion”.

This precision also applies to the level of detail. Indeed, “explain it to me like a beginner” and “answer as a domain expert” produce two radically different responses, both useful depending on your need.

Mistake 4: Forgetting to Assign a Role

The same model can answer like a cautious lawyer, a creative marketer or a meticulous financial controller. It all depends on the role you assign. Without guidance, it adopts an average tone, neither truly expert nor truly tailored.

So start your important prompts with a role assignment: “You are an HR specialist who knows Swiss employment law” or “Act as a demanding proofreader tasked with finding the weaknesses in this text”. That single sentence steers the vocabulary, the technical depth and the analytical angle. In practice, the role works even better combined with the context from mistake 1: who the model is, who you are, what you expect. You then get answers that sound like a competent colleague, not an encyclopedia page.

Mistake 5: Accepting the First Answer

A model’s first answer is a draft, not a finished product. Yet many users copy it as is, then wonder why the output feels mediocre. Advanced users, on the other hand, iterate systematically.

Three follow-ups are often enough to transform a result. First ask for a critique: “What are the weaknesses of your answer?” Then demand alternatives: “Propose three versions with different angles.” Finally tighten: “Keep version 2, cut it in half and make it more concrete.” Each round refines the result, exactly as it would with a colleague. Furthermore, if the answer heads in the wrong direction, do not patch it up: rewrite the initial prompt with the missing details. Starting clean costs less than fixing a wobbly text.

Mistake 6: Sharing Confidential Data

Pasting your client list, a named contract or salary data into a consumer tool: that is the mistake that can cost far more than a bad text. Indeed, depending on the tool’s configuration, your data may feed model training or transit outside Switzerland. The revised Swiss Data Protection Act strictly governs such processing. And the risk is far from theoretical: the Qualinsight study reported by PME.ch shows that a large share of workers in French-speaking Switzerland use AI secretly, outside any framework.

Adopt three simple reflexes. First, anonymise: replace names, amounts and identifiers with variables before pasting a document. Second, use the professional versions of the tools, which exclude training by default. Finally, follow your company’s policy on approved data. If that policy does not exist yet, that is a signal: governance needs to catch up with usage. A perfect prompt is worthless if it exposes your company legally.

Mistake 7: Never Checking the Facts

Generative AI models sometimes produce false statements with perfect confidence: invented figures, non-existent sources, imaginary laws. This hallucination phenomenon is shrinking with newer models, but it has not disappeared. The vendors’ own guides all repeat it: human verification remains essential.

The rule is simple: every checkable fact must be checked before publication. Dates, amounts, quotes, legal references, statistics. Ask the model for its sources, then actually verify them, because a cited source is not always an accurate one. For sensitive topics, cross-check with a classic search. This discipline turns AI into a reliable assistant: it speeds up production while you keep responsibility for the final content. Power from the tool, judgement from the human.

Conclusion

These seven mistakes share one trait: fixing them requires no technical skill whatsoever. Provide context, break down your requests, specify the format, assign a role, iterate, protect your data and check the facts: anyone can apply these rules from their very next prompt. Results follow immediately, and the gap widens fast between those who master these basics and those who improvise. Prompt engineering is no longer a specialist craft anyway: it is becoming a basic workplace skill, much like Excel twenty years ago. Better to build it deliberately than to leave it to chance. To go further, our artificial intelligence resources and training courses cover these techniques with exercises on your real business cases.

FAQ

What is prompt engineering?
It is the craft of structuring your instructions to a generative AI to get reliable, usable answers: context, role, clear instruction, output format and iteration.

What is the most common prompting mistake?
Lack of context. A request with no situation, goal or audience produces a generic answer, whatever the tool.

Does a good prompt work across all tools?
Mostly, yes. Context, role, format and iteration improve results in ChatGPT, Claude, Gemini and Copilot alike. Only fine-tuning details vary between models.

Do you need technical skills to improve at prompting?
No. The seven mistakes in this article can be fixed without a single line of code. Progress comes from structured practice, ideally on your own business cases.

Can you share internal documents with ChatGPT?
Only within your company’s rules. Anonymise sensitive data and favour the professional versions of the tools, which exclude training by default.

How do you know whether an AI answer contains errors?
Check the facts systematically: dates, figures, sources and references. Ask the model for its sources and verify them yourself before publishing anything.