Polite prompts cause AI errors: how to write boring prompts

When I started with prompt engineering, I thought it meant being a prompt whisperer, someone who knows the magic words to charm the AI. An MIT study, new when I wrote this in December 2025, shows why that idea causes errors.

Language models follow the shape of a sentence more than its logic. A polite, complex prompt gives the model more grammatical patterns to copy, so it can ignore your real instruction and hallucinate. The fix is a boring prompt: flat, structured, with no politeness, written as labeled fields, lists and XML tags.

What did the MIT study find about AI and sentence structure?

The MIT study found that a model can answer from the grammatical pattern of a question, without understanding the question. In the paper, the researchers trained models where specific topics were linked to specific sentence structures.

The researchers give this example. A model learns that "Where is Paris located?" gets the answer "France", which is correct. Then it gets a nonsense sentence with the same grammatical structure (adverb + verb + proper noun + verb): "Quickly sit Paris clouded?" The model can answer "France" again.

The model did not understand the geography. It understood the syntax. Syntax is the structure of a sentence: the order and type of its words. The model recognized the shape of the sentence and guessed the topic linked to that shape. This happens because an AI model generates the most probable output from the input and from everything it has generated so far.

Why polite prompts make AI hallucinate

Polite prompts add syntactic noise. A hallucination is an AI output that sounds right but is not grounded in the input. When you write complex, polite, professional letters to your AI agent, you give the model more grammatical patterns to latch onto. The model then has a higher chance to ignore your instructions and hallucinate from the tone of your sentence structure.

A real example: the delivery date that kept going wrong

I reviewed a document extraction workflow. The workflow had to extract the "Delivery Date" from messy PDFs. It kept returning the billing date from the invoice header instead of the date in the shipping body.

This was the original prompt, which failed:

"Could you please review the attached document and identify the date when the shipment is expected to arrive? Please let me know if you're unsure."

My hypothesis: the polite, conversational structure triggered a generic "document summarization" pattern in the model, not a specific extraction task. So I applied the boring protocol. I removed all politeness, all complex clauses and all humanity.

This is the new prompt, which works:

"Task: Extract delivery date. Context: Shipping document. Output format: YYYY-MM-DD. Constraint: Ignore billing date."

The result was 100% accuracy on 10 test outputs.

How to write boring prompts in 3 steps

If an automation works 80% of the time but fails randomly, the cause is probably syntax. Do these three steps in order. (Random failures can also come from bad input data, which I cover in why your data, not your AI, is the problem.)

1. The caveman audit. Read each system prompt. A system prompt is the standing instruction the model gets before every task. Remove these words and phrases:

   • "Please", "Kindly" and "I would appreciate".
   • Complex transition words, such as "Furthermore", "Consequently" and "In regards to".
   • Polite padding, such as "Hope you are well".

   These words cost money, because each one uses tokens (the units a model reads and bills by). They also distract the model with patterns it can copy.

2. Break the prose. The MIT study found that grammatical patterns trigger hallucinations. The best way to break a grammatical pattern is to stop using grammar. Write lists, not paragraphs. Use XML tags (labels in angle brackets, such as <input>) to separate instructions from data. The model then processes the logical structure of the task, not the linguistic flow.

   • Bad: "Read the email below and if the user is angry draft a response that apologizes but don't promise a refund."
   • Good (XML): put the email in <input> tags. Then add a <rules> block: 1. Analyze sentiment. 2. If sentiment = negative, draft apology. 3. Constraint: No refunds.

3. The meta-rewrite. Let an LLM do the rewrite for you. A meta-prompt is a prompt that tells a model how to rewrite another prompt. I use this meta-prompt to clean up client instructions before they go into production:

   "Rewrite the following prompt to be grammatically flat and boring. Remove all politeness, conversational filler, and complex sentence structures. Convert instructions into a bulleted list or XML structure. Prioritize direct logic over linguistic flow. [Insert your current prompt]"

A boring prompt is one part of a reliable automation. The other parts are the input checks and output checks around the model, which I describe in 6 patterns for a bullet-proof automation.

Polite prompt vs boring prompt

The bottom line

The MIT study proves that prompt engineering is not about charm. The best prompt engineer writes like a drill sergeant: clear, direct and boring. If your AI hallucinates, stop trying to be Shakespeare and start being Sheldon Cooper.

Build a calmer business,Yuri

Frequently asked questions

Why does a polite, conversational prompt make AI give wrong answers?

An MIT study found that language models link sentence structure to topics and can answer from the shape of a sentence, not its meaning. A polite, complex prompt adds more grammatical patterns for the model to follow. That syntactic noise increases the chance the model ignores the real instruction and hallucinates an answer that fits the tone of the prompt.

What is a boring prompt in prompt engineering?

A boring prompt is a flat, structured instruction with no politeness, filler or complex clauses. It names the task, the context, the output format and the constraints in separate labeled lines, for example: Task, Context, Output format, Constraint. Yuri Vonchitzki of Ninjabot used this form to fix a delivery-date extraction prompt, which then gave 100% accuracy on 10 test outputs.

How do I fix an AI automation that works 80% of the time but fails randomly?

Random failures in an AI automation are often a prompt syntax problem. Do three things. First, remove polite words, transition words and padding from the system prompt. Second, replace paragraphs with lists and use XML tags to separate instructions from data. Third, use an LLM with a meta-prompt to rewrite your prompts into a flat, bulleted or XML structure before production.

Why use XML tags in an AI prompt?

XML tags separate the instructions in a prompt from the data the model must process. For example, put a customer email inside input tags and the steps inside a rules block. Lists and tags break grammatical patterns, so the model follows the logical structure of the task and not the linguistic flow of the sentence.