How the Prompt Engineering Mistakes Quiz Works
The Prompt Engineering Mistakes Quiz tests your ability to identify common problems in real-world prompts. Each question presents a system prompt or user prompt with a specific issue, and you must identify what is wrong from four options. After answering, you see a detailed explanation of the mistake, why it causes problems in practice, and a before-and-after comparison showing how to fix it. The quiz covers 15 categories of mistakes that account for the vast majority of quality issues in production LLM applications.
The quiz is designed for both beginners learning prompt engineering fundamentals and experienced practitioners who want to validate their knowledge. The questions progress from obvious issues like vague instructions to subtle problems like instruction ordering effects and implicit assumption errors. Each question is based on mistakes observed in real production prompts, not theoretical edge cases. The fix suggestions are practical and immediately applicable to your own prompts.
The 15 Categories of Prompt Engineering Mistakes
Vagueness is the most common mistake and the easiest to fix. Prompts that say "be helpful" or "answer well" provide no actionable guidance to the model. The model interprets vague instructions differently on each call, producing inconsistent outputs. The fix is specificity: replace every vague adjective with a concrete instruction. "Be concise" becomes "answer in 2-3 sentences." "Be accurate" becomes "cite the source document for every claim." Specificity is the single highest-leverage improvement for any prompt.
Missing output format specification causes the second most production issues. When the prompt does not specify the expected output structure, the model chooses its own format, which varies between calls. One response might be a paragraph, the next a bullet list, the next a numbered list. Downstream code that parses the output breaks unpredictably. Always specify the exact output format: "Respond with a JSON object containing keys: summary (string), sentiment (positive/negative/neutral), confidence (float 0-1)."
Token waste occurs when system prompts include unnecessary context, redundant instructions, or excessive politeness that consumes tokens without changing model behavior. "You are an incredibly helpful, knowledgeable, and friendly assistant who always aims to provide the best possible answer to every question" can be compressed to "You are a helpful assistant." The model already tries to be helpful, so restating it wastes 15 tokens per request. At 10,000 requests per day, that is 150,000 wasted tokens daily, costing $0.45 per day with Sonnet pricing. Multiply this across every unnecessary token and the waste becomes significant.
Hallucination risk increases when prompts encourage the model to answer questions it does not have information for. Prompts that do not include an "I don't know" escape hatch force the model to generate an answer even when it is uncertain. The fix is explicit uncertainty handling: "If the answer is not in the provided documents, say 'This information is not available in the provided context.' Do not guess or infer beyond what is explicitly stated." This instruction is the single most effective hallucination reduction technique and should be included in every factual system prompt.
Instruction ordering effects are a subtle mistake. Models pay more attention to instructions at the beginning and end of the system prompt than instructions in the middle. Burying the most important constraint in the middle of a long prompt reduces its effectiveness. Place your most critical instructions first and last. Format rules, safety constraints, and output specifications should be at the top. Secondary preferences and style guidelines can go in the middle.
Common Patterns in Bad Prompts
The kitchen sink prompt tries to handle every possible scenario in a single system prompt. It contains dozens of instructions covering every edge case the developer thought of, resulting in a 3,000-token prompt that the model struggles to follow consistently. The fix is to prioritize: identify the 5 most important instructions and cut everything else. If specific edge cases need handling, add them as conditional rules rather than blanket instructions.
The conflicting instruction prompt gives the model contradictory directions. "Be concise" followed by "provide thorough explanations" three paragraphs later creates a conflict the model resolves unpredictably. Review your prompt for contradictions by reading each instruction in isolation and checking whether any pair conflicts. When you need both conciseness and thoroughness, be explicit about when each applies: "For simple questions, answer in 1-2 sentences. For complex questions, provide a detailed explanation."
The role confusion prompt assigns a role that conflicts with the task. "You are a creative storyteller" combined with "extract the exact dollar amounts from this invoice" confuses the model about whether to prioritize creative interpretation or exact extraction. The role should match the primary task. For extraction tasks, use roles like "precise data extractor." For creative tasks, use roles like "creative writer." Mismatched roles degrade output quality on the primary task.
Privacy and Local Execution
The Prompt Engineering Mistakes Quiz runs entirely in your browser. Quiz questions, answers, scores, and progress are processed client-side using JavaScript. No data is sent to any server. There are no accounts, no cookies, no analytics, and no server-side processing. Your quiz results remain completely private on your device.