> ## Documentation Index
> Fetch the complete documentation index at: https://simpai.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.

# Common prompting mistakes that cause hallucination

> These are the most frequent prompting errors S.I.M.P. detects — from missing context to ambiguous scope — and how to fix each one before running your prompt.

Most hallucination-prone prompts share the same handful of problems. S.I.M.P. detects these issues automatically, but understanding what they are — and why they matter — helps you avoid them before you ever hit submit. Below are the mistakes that appear most frequently in high-risk prompts, and how to fix each one.

<AccordionGroup>
  <Accordion title="No context about who you are or what this is for">
    When you don't tell the model who is asking or why, it invents an audience. It might assume you're a student, an expert, a business, or a casual user — and tailor the response to that assumption rather than to your actual needs.

    **Fix:** Start your prompt with a brief framing sentence. "I am a nurse explaining medication instructions to a patient" gives the model a concrete target. "I am preparing a board presentation for a non-technical audience" tells the model what level of detail and vocabulary to use.
  </Accordion>

  <Accordion title="Open-ended scope">
    "Tell me about X" is one of the most common patterns in high-risk prompts. The model has to decide on your behalf what aspect of X to cover, how much detail to include, and where to stop. Those decisions are almost never what you had in mind.

    **Fix:** Replace open-ended phrasing with a defined scope. Instead of "tell me about X," try "give me a 3-point summary of X's key risks" or "explain how X works in under 150 words, for someone who has never heard of it."
  </Accordion>

  <Accordion title="Implicit assumptions">
    You know what you mean. The model doesn't. When your prompt relies on background knowledge that isn't written down, the model either ignores it or guesses at it — and a wrong guess can silently derail the entire response.

    **Fix:** Make your assumptions explicit. If you're asking about a specific product, name the version. If you're asking about a process, describe the relevant step you're already at. If there's industry-specific terminology the model might interpret differently, define it.
  </Accordion>

  <Accordion title="No output format specified">
    Without format guidance, the model picks whatever structure it considers appropriate for the topic. This may be a long narrative paragraph when you needed a table, or a bullet list when you needed prose, or a technical breakdown when you needed a plain-language summary.

    **Fix:** Specify the format explicitly. Tell the model to respond as a table, a numbered list, a JSON object, a short paragraph, or whatever format you'll actually use. If length matters, state a word or sentence count.
  </Accordion>

  <Accordion title="Asking multiple questions in one prompt">
    When a prompt contains several questions, the model may answer some and skip others — often without making it obvious which ones it addressed. It may also blend answers together in a way that makes each one harder to evaluate individually.

    **Fix:** Break multi-question prompts into separate, focused prompts. If you need all the answers in one response, number each question explicitly and instruct the model to answer them in order.
  </Accordion>

  <Accordion title="Using vague qualifiers like &#x22;good&#x22;, &#x22;recent&#x22;, or &#x22;best&#x22;">
    These words mean different things in different contexts, and the model will apply whatever interpretation is most statistically common — which may not be yours. "Recent" could mean last week or last decade. "Best" could mean most popular, most technically rigorous, or most beginner-friendly.

    **Fix:** Replace vague qualifiers with specific criteria. Instead of "recent research," say "research published after 2022." Instead of "best practices," say "practices recommended by OWASP for web application security."
  </Accordion>

  <Accordion title="No constraints on length or depth">
    An unconstrained prompt often produces a response that is either too brief to be useful or padded with filler to seem thorough. The model optimizes for appearing helpful, not for matching the depth you actually need.

    **Fix:** Set explicit length and depth constraints. Specify a word count, a number of bullet points, or a level of detail ("suitable for a technical audience" vs "suitable for a general audience"). If you want the model to go deep on one specific aspect, name that aspect.
  </Accordion>
</AccordionGroup>

<Warning>
  A Critical risk score (9–10) usually means your prompt is missing fundamental context. The AI has almost nothing to anchor its response to.
</Warning>

After S.I.M.P. analyzes your prompt, the changes panel lists each modification made to the sanitized version and explains why it was necessary. Use that panel to understand exactly which of these mistakes appeared in your prompt — not just to fix the current prompt, but to recognize the pattern in future ones.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.