There’s a counterintuitive trap waiting for anyone who gets serious about prompting AI models: the more effort you put into your prompt, the worse your results can get. Not always, not inevitably, but often enough that it’s worth understanding the mechanism.
This isn’t a reason to stop thinking carefully about how you communicate with AI. It’s a reason to think differently about it.
1. Over-Specified Prompts Collapse the Model’s Reasoning Space
When you pile constraints into a prompt, you’re not just guiding the model. You’re narrowing the set of valid completions it can generate. The model’s job becomes satisfying your constraints rather than actually solving your problem. These are not the same thing.
Ask a model to “write a three-paragraph summary, using only simple language, avoiding passive voice, starting with a statistic, and targeting a CMO audience” and you’ll get something that checks all those boxes while completely missing the point of whatever you wanted summarized. The model optimizes for constraint-satisfaction. You wanted insight.
The fix is to separate your constraints from your goal. State the goal clearly first. Add one or two constraints that are genuinely load-bearing. Drop the rest.
2. Chain-of-Thought Prompts Can Produce Confident Wrong Reasoning
The “think step by step” technique genuinely improves performance on many tasks, particularly math and logic problems. But it has a less-discussed failure mode: models will produce fluent, structured chains of reasoning that are wrong at every step, and present them with complete confidence.
When you prompt for explicit reasoning, you get explicit reasoning. Whether that reasoning is sound is a separate question. The model has no internal audit process that checks its work before outputting it. You’re reading a transcript of the model generating text that looks like careful thinking, not evidence that careful thinking occurred. As this piece on confident hallucination covers, capability and accuracy aren’t the same thing.
Use chain-of-thought prompting for genuinely structured problems where you can verify the steps. Don’t use it for open-ended creative or analytical tasks and then trust the reasoning because it came with numbered steps.
3. Few-Shot Examples Anchor the Model to Your Examples, Not Your Intent
Few-shot prompting (providing examples of the output you want before asking for the real thing) is one of the most reliable techniques in the toolkit. It also carries a specific failure mode that catches people off guard.
The model learns the surface pattern of your examples, not the underlying principle you think you’re demonstrating. Give three examples of concise product descriptions and the model may match their length, tone, and sentence structure without understanding why those things made the examples good. Change any feature of the real task and the outputs often degrade in ways that feel random but aren’t: the model is still pattern-matching to your examples, which no longer fit.
When you use few-shot examples, vary them enough that no single surface feature (length, format, vocabulary level) becomes the implicit rule. Or explain explicitly what makes each example good, rather than letting the model reverse-engineer it.
4. Role Prompts Often Produce Stereotype, Not Expertise
“You are an expert startup attorney” or “you are a senior data scientist” sounds like a smart way to calibrate the model’s responses. In practice, the model’s representation of “expert startup attorney” is built from whatever text about startup attorneys exists in its training data, which skews heavily toward general descriptions, listicles, and blog posts written about lawyers, not by them.
You may get outputs that pattern-match to how people talk about expert behavior rather than how experts actually reason. The framing feels rigorous but the outputs often read like someone doing an impression.
The better move is to describe the task context rather than the persona. Instead of “you are an expert contract negotiator,” try “I’m evaluating this SaaS contract clause for a small company with no legal team. Flag anything that creates meaningful risk and explain why.” You’ve given the model the thing it actually needs: situational specificity.
5. Long Context Prompts Bury Your Actual Question
Models handle long prompts worse than the benchmarks suggest. Research on “lost in the middle” effects (work from Stanford and others published in 2023) found that language models perform significantly worse at retrieving information placed in the middle of long contexts compared to information at the beginning or end. You can write a detailed, nuanced 2,000-word prompt and have the model effectively ignore the most critical instruction you buried in paragraph four.
This matters for anyone building complex multi-step prompts or feeding in long documents with instructions. Your careful elaboration may literally be getting downweighted by the model’s attention mechanisms.
Put your most important instruction last, not first. Keep system prompts tight. If you need the model to work with long source material, consider chunking it rather than dumping everything in one shot.
6. Prompt Engineering Fluency Can Make You Worse at Noticing Bad Output
This is the most uncomfortable one. When you’ve spent time learning to prompt well, you start reading outputs differently. You know what good outputs look like. You’ve trained yourself to recognize the patterns. This same skill makes it easier to mistake a fluent, well-structured wrong answer for a correct one.
The model has also, in a real sense, learned to produce outputs that satisfy people who write careful prompts. It’s very good at sounding like it understood what you asked. The prompt you wrote isn’t always the prompt the model read, and the more elaborate your prompting, the more places that gap can hide.
The discipline that actually protects you isn’t better prompting. It’s maintaining verification habits that don’t depend on how good the output looks. Check claims. Test outputs against cases you know the answer to. Don’t let fluency substitute for correctness.
The Actual Principle
Better prompting isn’t about adding more. It’s about being precise about what matters and ruthlessly dropping everything else. The models that respond best to simple, clear, direct requests aren’t unsophisticated. They’re reflecting something true about communication: clarity transfers intent, and elaboration often obscures it.
