Common Prompt Mistakes and Fixes

Six Mistakes That Quietly Sabotage Your Results

When a language model returns something strange, most people assume the model failed. That instinct is usually wrong. More often the output is the natural consequence of a prompt that is vague, ambiguous, or contradictory. The good news is that the same six problems appear again and again, and each has a small, reliable fix. Learn to recognize these errors and you will catch them in your own prompts before they cost you a frustrating round. Here are the six most common prompt mistakes and the edits that solve them.

Mistake One: Vague Goals

The prompt "help me write a summary" gives the model almost no information. A summary of what, at what length, for which audience, in what format? Every unanswered question becomes a guess, and you may not like how it guesses.

# Vague
Summarize this.

# Specific
Summarize the attached article in exactly 100 words for a busy
product manager. Lead with the single most important insight,
then list three supporting points as bullets.

The fix: answer five questions before you prompt: what, why, for whom, in what format, and how long. If you cannot answer them, the model cannot either.

Mistake Two: Overloaded Instructions

Stacking too many requirements in one prompt produces a muddle. The model juggles format, tone, length, examples, and content all at once, and something drops. The longer the instruction list, the higher the chance of a conflict or an overlooked clause.

• Write the main instruction as its own sentence.
• Move secondary constraints (tone, format) into a short bullet list.
• If the task is large, split it across multiple turns.

The fix: trim. Aim for one clear goal plus a handful of constraints. If you have more than five separate requirements, suspect overloading and reduce.

Mistake Three: Contradictions

Prompts sometimes ask for two things at once that pull in opposite directions. "Be brief but explain everything in detail" is a classic. So is "sound casual, but use formal terminology." When instructions conflict, the model has to pick a compromise, and the result satisfies no one.

Contradictions force the model to choose which rule matters more, and it will not tell you which one it picked. Scan your prompt for opposing pairs: brief vs. detailed, fun vs. serious, concise vs. comprehensive. Resolve each pair before sending.

The fix: decide priorities. If brevity matters more than detail, say "prefer short sentences; sacrifice depth before clarity." Declaring the tiebreaker removes the ambiguity.

Mistake Four: Zero Constraints on Format

Models default to paragraphs because essays are common in training data. If you want a table, a bullet list, a numbered checklist, or a one-sentence answer, you must say so. Unspecified format is the reason so many outputs look like interchangeable term papers.

# No format requested
List the pros and cons.

# Format specified
Present the pros and cons as a markdown table with three columns:
category, point, and why it matters.

The fix: always state the desired output shape. When the format is explicit, both you and the model spend less energy guessing.

Mistake Five: Expecting Mind Reading

Prompts that refer to "the project," "our customers," or "the attached file" without further context assume the model can read your mind. Unless you paste the material or summarize it, the model improvises from general knowledge, which may not match your reality.

The fix: supply context inline. Paste the relevant text, give a one-paragraph briefing on the situation, or name the key facts the model should assume. The cost of a few extra sentences of context is trivial compared with the cost of a wrong answer built on false assumptions.

A useful shortcut is the "briefing sentence." Before your real question, add one line that places the model in your world, such as "Context: we sell a calendar tool and users are non-technical." It costs ten words and removes an entire class of wrong guesses. Every detail you omit, the model fills with its own assumption about what you meant.

Mistake Six: Ignoring the Output

The last mistake is forgetting that the output is feedback. When the model misinterprets your prompt, the useful response is not frustration; it is revision. The gap between what you wanted and what you got is a direct signal about what your instructions were missing. Each failed output is a diagnostic clue.

• If output is generic, add specifics and examples.
• If output is wrong, add boundaries and exclusions.
• If output is wordy, add length and style constraints.
• If output drifts, add a system-style standing rule.

The fix: treat every output as a prompt review. Ask "what does this tell me about my instructions?" and adjust. Iteration converts failures into improvements.

A Deeper Look at Vague Language

Vague adjectives are another silent saboteur. Words like "good," "clear," "effective," and "friendly" carry no shared definition, so the model guesses at a meaning. Replace them with measurable descriptions whenever you can. Instead of asking for a "good summary," ask for a summary that "misses no key numbers and uses under 100 words." Instead of "friendly tone," say "use contractions, address the reader directly, and avoid passive voice." Each concrete substitute removes one more guess the model would otherwise make, and each guess you remove is a point where your result improves.

It also helps to name the failure mode you want to avoid. If you are worried the model will invent facts, say "do not add anything not present in the source." If you worry it will be too polite to criticize, say "do not soften your critique." Telling the model what not to do, explicitly and specifically, is often as valuable as telling it what to produce. Too many prompts phrase only the positive goal, leaving the model free to fail in a handful of predictable ways that a single negative clause would have prevented.

A Checklist Before You Press Send

Run this quick mental checklist on any important prompt. It catches most of the six mistakes before the model ever sees your text.

• Do I know exactly what I want as the result?
• Is the main instruction a single, clear sentence?
• Do any requirements contradict each other?
• Did I specify the output format and length?
• Did I provide enough context, or am I assuming the model knows?
• Am I ready to read the output as feedback and adjust?

The Bottom Line

None of these mistakes reflect a broken model. They reflect prompts that captured too little of what the writer knew. By making goals specific, cutting overload, resolving contradictions, demanding a format, supplying context, and treating outputs as feedback, you fix the vast majority of disappointing results. The edits are small, but they compound: a clean prompt produces a useful draft, which saves you a whole cycle of correction. Master these six fixes and your prompting will feel less like gambling and more like engineering.