The 10 Most Common Prompting Mistakes

You know the techniques. You understand the principles. Yet your results remain inconsistent. Often, the culprit isn’t a technical gap but an ingrained habit—invisible precisely because you repeat it daily. The ten mistakes below plague nearly every regular user, even experienced ones. The table summarizes them; the rest of this lesson details the most costly ones.

The vague prompt, and its two cousins

The first mistake is the most widespread and the most costly. “Help me with my presentation” contains none of the information the model needs: no audience, duration, objective, or format. It will therefore produce a response plausible for any presentation, meaning useful for none. The fix is simple: write down what you had in mind but didn’t say.

Create a detailed outline for a 20-minute presentation to
the executive committee. Topic: Q1 review and Q2 outlook
for the marketing team. Include key KPIs, wins,
losses, and 3 recommendations for the next quarter.
Format: One slide per point, with a title and 3-4 bullets max.

The second mistake is the opposite extreme: asking for everything at once. A prompt like “Write a blog post, create the images, write promotional posts for LinkedIn and Twitter, do the meta description, and propose a publishing schedule” won’t yield five good deliverables. It yields five drafts, none of which are usable as-is. Handle each element in a separate prompt, reusing the previous result as context for the next.

The third concerns output. Without a format instruction, “Analyze these sales data” produces free-form text that you must reorganize before sharing—a preventable time sink that you pay for with every analysis.

Analyze these sales data and produce:
1. An executive summary (5 lines)
2. A table: Product | Revenue | YoY Change | Trend
3. The top 3 alerts
4. The 2 identified opportunities

Blind trust and giving up too soon

ChatGPT states falsehoods with the same confidence as truths, especially regarding figures, dates, quotes, and proper nouns—exactly the elements a busy reader won’t verify. Adding a self-reporting instruction systematically at the end of your prompt costs nothing and tells you where to focus your verification efforts.

State your confidence level for each assertion
(certain / probable / uncertain).
If you are unsure about a piece of information, state it clearly
rather than inventing it.

The symmetric mistake is closing the conversation after a disappointing response, concluding the tool can’t do the job. Most often, the prompt needs refining, and precise feedback achieves this in one or two iterations: stating what’s wrong, point by point, works better than reformulating the request from scratch.

Your response doesn't work for me for these reasons:
1. The tone is too formal—I want something more conversational
2. You forgot to mention the budget
3. The examples come from large corporations—I need
   examples suited to a 30-person SME

Rewrite correcting these 3 points.

The missing role and contradictory instructions

Without an assigned role, the model responds as a generalist assistant: correct, flat, and lacking angle. With a specific, contextualized role, it mobilizes industry vocabulary, references, and priorities. The difference shows in the first sentence—a CFO and a CMO won’t comment on the same dashboard in the same way.

Conversely, an overloaded prompt ends up contradicting itself. “Be concise but detail every point” or “use a professional but casual tone” forces the model into an arbitrary choice. You’ll get one or the other at random, without understanding why. So, review your constraints and check their compatibility. When two requirements truly conflict, prioritize rather than choose: “Prioritize conciseness. If a point deserves expansion, limit it to 3 lines.”

Waste: disposable prompts and borrowed prompts

You write an excellent prompt for a meeting report, then lose it. The next week, you rewrite a weaker one, and so on. You can fix this silent erosion by saving your best formulations as templates with variables in brackets. Conversely, prompts found online were designed for a context that isn’t yours: use them as inspiration, then replace the role, context, and constraints with your own before running them. Otherwise, you’ll get a response perfect for someone else’s company.

The wrong tool for the job

The last mistake is more strategic than technical: using ChatGPT the same way for everything. It excels at writing, analysis, structuring, and brainstorming. For calculations, ask it to use data analysis, which performs calculations in Python, and verify the result. For recent information, rely on web search and its sources. And for decisions requiring human judgment, it can prepare the brief, but the decision remains yours. Knowing which tool to mobilize will save you as much time as knowing how to write a prompt.

Your turn

Audit your last five conversations. Reopen your last five exchanges with ChatGPT and run each through the table filter: which mistake from the list did you make, and how would you have phrased the request to avoid it? Rewrite the corrected prompt, run it again, and compare the two responses. The improvement is immediate, but the main lesson lies elsewhere: this audit reveals which of these ten habits is truly yours, and that is the only one you need to work on.

Official sources: Prompting (ChatGPT documentation) · How do I create a good prompt for an AI model? · Does ChatGPT tell the truth? · Data analysis with ChatGPT (OpenAI Help Center)

Key takeaways

• Precision is the antidote to most mistakes: goal, context, output, and limits
• Break down complex tasks instead of asking for everything at once
• Specify the output format to get a directly usable result
• Don’t trust blindly: have uncertainties flagged and verify sources
• Mobilize the right tool: data analysis for calculations, web search for news