10 beginner mistakes that ruin your prompts

"My code doesn't work, fix it." A developer in a hurry who sends that without pasting the code or the error will still get an answer, long and confident, with a list of possible causes. AI always answers, and that's what makes bad prompts hard to spot: nothing tells you something was missing.

Here are ten common mistakes, grouped by type, with the fix to apply for each.

Four mistakes in how you word the request

1. Keeping the goal to yourself

"Summarize this report" doesn't say why. A summary meant to prepare a budget meeting doesn't keep the same passages as a summary meant to brief new hires. Add a sentence about the use: "to decide in Monday's meeting whether we renew the contract."

2. Asking for everything in one message

A garage manager who asks in one go for an analysis of his customer reviews, an action plan, a template reply, and a social media post will get four botched jobs. Separate the steps and check each one before the next. The method is described in our article on prompt chains.

3. Sticking to adjectives

"A professional and engaging text" means almost nothing, because everyone reads into it whatever they like. Describe what shows on the page instead. Here's the before:

Write a professional and engaging introduction for my accounting firm.

And here's the after, where each adjective is replaced by a checkable instruction:

Write the introduction for my accounting firm for the homepage of my website.

Readers: craftspeople and shopkeepers in [CITY] who are looking for their first accountant.
Information: [YEAR FOUNDED], [TEAM SIZE], [SERVICES], [WHAT SETS US APART].

Write in the first person plural, with sentences of 20 words maximum and no accounting term left unexplained. Start with the reader's concrete problem, not with our history. 80 words maximum.

4. Saying only what not to do

"Don't be too long, not too formal, no jargon" leaves the AI without direction. A positive instruction steers it better: "three sentences, as if to a curious neighbor who doesn't know the trade." Save prohibitions for what's truly off the table, like a word to ban.

Three mistakes in what you provide

5. Believing the AI knows your situation

It doesn't know your client is angry or that your shop closes in August, unless you wrote it in the conversation or saved it in the tool's memory. Write what a stand-in would need to know to do the task. Our article on context helps you sort it out.

6. Asking it to analyze a document without providing it

"Summarize the latest report from such-and-such organization" or "what does this contract say?" with no attachment or pasted text invites the AI to improvise. Depending on the tool, it may run a search, or write something from memory that's plausible and wrong. Paste the text, or attach the file if the tool allows it, and ask it to rely only on that.

7. Giving no expected format

Without a format, you get an introduction, generalities, and a closing offer of help. Say what you want: "a three-column table," "five bullet points," "two paragraphs with no introduction." We go into these settings here.

Three mistakes in how you work

8. Throwing everything out instead of fixing it

An average answer often contains a useful part. Say what works and what doesn't: "keep the structure, the tone is too salesy, cut the last sentence." Conversely, when a conversation has bogged down on a wrong track, start over in a new conversation with a rewritten prompt rather than piling up corrections.

9. Trusting the answer because it's well written

A reference, a figure, or a legal article can be made up and still look flawless. Before reusing a fact, find it in a source. A craftsman who pastes a legal notice into his quote because the AI suggested it is still responsible for what he sends. Get into the habit of checking answers methodically.

10. Pasting data that shouldn't leave your hands

Payslips, medical records, a customer number, a confidential contract: depending on the tool, the plan, and your settings, what you send may be stored or used for training. Replace names with "Client A" and remove whatever identifies someone. Read the tool's terms, and if you handle personal data at work, check your organization's rules. The CNIL publishes recommendations on the subject.

Have your prompt audited before you send it

You can ask the AI to spot these mistakes. The following prompt works well for a task you care about.

Here is a prompt I'm about to use:

"""
[PASTE YOUR PROMPT]
"""

Don't run this prompt. Act as a proofreader and answer only these questions:

1. What information are you missing to do this job properly?
2. Which words could be understood in more than one way?
3. What decisions would you be forced to make in my place?
4. Are there several tasks that should be separated?

Then suggest a corrected version, with [BRACKETS] where I need to add information you don't have.

This proofreader has its limits: it can ask for needless details or lengthen the prompt for no reason. Keep what seems right to you and ignore the rest.

Where to start

Fixing all ten at once makes no sense. Go to your history, open your last three prompts, and look for the mistake that keeps coming back. Fix that one for a week, on every prompt, before moving on to another.