Skip to content
Tutorials Center
My learning

Topics

Lesson 10 of 38 ยท Prompting Fundamentals

Common Prompt Mistakes and How to Fix Them

Ten frequent prompt-writing mistakes, from vague tasks and buried instructions to conflicting rules and fake certainty, each with a fix and a runnable checker.

Most bad outputs trace back to a small set of repeat mistakes. Learning to spot them is like learning common bugs in a programming language: once you can name them, you find them quickly. This lesson lists ten, with a before and after for each, then gives you a small checker you can extend.

1. A vague task

Before: "Look at this and tell me what you think." After: "Review this function for bugs. List each bug with the line number and a one-line fix."

Fix: name the verb (review, classify, rewrite) and the deliverable.

2. Missing audience or purpose

Before: "Explain Kubernetes." After: "Explain Kubernetes in 150 words to a manager who will decide whether the team should adopt it. Focus on cost and risk."

3. Stacking too many jobs

One prompt that asks for translating, summarizing, extracting dates, scoring sentiment, and writing a reply will do each part less well. Split it into steps, each with its own prompt and check. See prompt chaining.

4. Conflicting instructions

"Be extremely detailed but keep it under 30 words." The model cannot satisfy both, so it picks. Read your prompt for contradictions and decide which rule wins ("If they conflict, the word limit wins").

5. Negative-only instructions

"Don't use jargon. Don't be long. Don't be rude." Lists of "don't" tell the model what to avoid but not what to aim for. Rewrite positively: "Use plain English a new customer would understand, in at most 3 sentences, with a polite tone."

6. Burying the key instruction

In a long prompt, the one critical rule hidden in the middle may get less weight. State it near the start, and repeat it at the end for long inputs.

7. No fallback for missing information

Without a rule for "I don't know", models guess. Add: "If the answer is not in the text, say 'Not stated.'"

8. Trusting the model with exact work

Counting words, adding long lists of numbers, sorting thousands of rows, or checking a date arithmetic puzzle is code's job. Let the model write or call code, or do it in your program.

9. Testing on one example

Covered in Iterating and Versioning Prompts. One passing example proves little.

10. Treating confident output as correct

Fluent text is not verified text. For anything factual or high-stakes, ground the answer in supplied documents, ask for the supporting quote, and verify programmatically where you can.

flowchart TB
  A["Bad output"] --> B{"Which mistake?"}
  B --> C["Vague task or audience"]
  B --> D["Conflicting or buried rules"]
  B --> E["Too many jobs"]
  B --> F["No fallback for unknown"]
  B --> G["Exact work done by the model"]

Diagnose before rewriting: each symptom points to a different fix.

A checker you can extend

The block below encodes four of these mistakes as simple heuristics. It is a teaching stand-in: real review needs a human, and every heuristic here has false alarms. Run it, then add your own rule.

Python Heuristic prompt reviewer

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

A debugging routine

When an output is wrong, run through these questions in order:

  1. Is the task stated? One sentence, one verb.
  2. Is everything needed present? Facts, audience, format.
  3. Do any rules conflict? Resolve with a priority.
  4. Is the fallback defined? What to do when unsure.
  5. Is the output cut off or random? Check settings.
  6. Is this really a prompt problem? Maybe split the task or use code.
Python A mistake-to-fix lookup

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

Check your understanding

Quick check ยท 5 questions

of 5 answered You scored / 5 ยท perfect!

  1. 1. Which prompt contains conflicting instructions?

    Detailed and under 30 words cannot both be satisfied, so decide which rule wins.
  2. 2. Why are lists of "don't" instructions weak?

    Positive, measurable instructions give the model something to aim for.
  3. 3. Which task is better done in code than by the model?

    Exact counting, sorting and arithmetic are code's job.
  4. 4. A prompt asks for translation, summary, date extraction and sentiment in one go. What is the best fix?

    Splitting tasks lets each step be simpler, testable and fixable on its own.
  5. 5. What does a fluent, confident answer prove?

    Fluency is not verification. Ground answers in supplied sources and check what you can.

Summary

  • Most failures come from a short list of mistakes: vague tasks, missing audience, too many jobs, conflicting rules, negative-only rules, buried instructions, no fallback, exact work given to the model, one-example testing and trusting fluent output.
  • Diagnose the symptom first, then apply the matching fix.
  • Cheap heuristic checkers catch obvious problems but never replace testing.

Next: Zero-Shot and Few-Shot Prompting