Common prompting mistakes: what the evidence says about personas, step-by-step prompts, formatting and context

Search for prompting advice and you will find a confident list of rules: give the model a persona, tell it to think step by step, offer a tip, say it is important, format everything in a particular way. Some of those rules were true of some models at some point. The question for anyone using AI at work in 2026 is which ones still hold, and what to do instead.

This article sorts the common prompting mistakes by what the evidence says about them. You will see where careful tests found little or no benefit from folk wisdom, where small changes really do swing results, and what the model makers themselves now recommend. It builds on the earlier articles in this series on how to prompt well and on context engineering.

The pattern: folklore is fragile, fundamentals are durable

Across the research reviewed here, one pattern repeats. Tricks that try to coax a model with a clever phrase show small, inconsistent or model-specific effects. Meanwhile, ordinary things like how clearly you describe the task, how much irrelevant material you include, and whether you ever test your prompt, show up as the larger and more reliable levers.

A few terms first. A prompt is the text you give a model. A benchmark is a fixed set of test questions used to score models. Accuracy here means the share of benchmark questions answered correctly. A reasoning model is one trained to work through a problem step by step before answering, so it often does so without being asked.

Mistake 1: Relying on magic phrases

Expert personas. A common tip is to open with "You are a world-class expert in…". A Wharton Generative AI Labs team tested this on six models, using two hard multiple-choice benchmarks (198 graduate-level science questions from GPQA Diamond and a 300-question subset of MMLU-Pro), with 25 responses per question per condition. Domain-matched expert personas produced no significant accuracy gains on GPQA. On MMLU-Pro two of the six models improved. For example, Gemini 2.0 Flash with an engineering-expert persona scored about 8.9 percentage points higher than with no persona. Meanwhile, a "toddler" persona lowered accuracy in four of the six models, and one model, Gemini 2.5 Flash, refused to answer an average of 10.56 of 25 attempts per question when given an unrelated expert persona³. The authors are careful to say these results are about factual accuracy only and that personas "may serve other purposes," such as setting tone³.

"Think step by step." Asking for visible reasoning, known as chain-of-thought prompting, helped earlier models. The same research group tested eight models on GPQA Diamond and found that for dedicated reasoning models the added benefit "appear[s] negligible"⁴. For non-reasoning models the average score rose by between 4.4 and 13.5 percentage points (the smallest gain was not statistically significant), but the picture was mixed on a stricter measure: requiring a question to be answered correctly every time, Gemini 2.0 Flash fell by 13.1 points and Gemini 1.5 Pro by 17.2 points⁴. Step-by-step prompting also cost time: 35% to 600% longer for non-reasoning models, 20% to 80% longer for reasoning models⁴. The authors note the limits of their work: one benchmark, a simple version of chain-of-thought, and a limited set of models⁴.

Threats, bribes and pleas. A third study tried prompts such as a $1,000 tip, a $1 trillion tip, a threat to punch, an appeal about a sick mother, and a shutdown threat, across five models and the same two benchmarks. Overall effects were small and inconsistent. The tip prompt, for instance, lowered one model's MMLU-Pro score by 5.5 points and raised another's by 5.3⁵. Yet on individual questions the swings could be large, up to 36 points better or 35 points worse, which is the reason some people swear by a trick: any single test can look like a win⁵. The authors conclude that the consistency of null results across models "provides reasonably strong evidence" that these tricks are not a dependable lever⁵.

The practical reading: these studies measured accuracy on exam-style questions with specific models in 2025. They do not show that every persona or reasoning prompt is useless, but they do show you should not assume one works.

Mistake 2: Assuming wording and layout are neutral

If phrases rarely help on average, you might conclude that wording does not matter. The evidence says otherwise, and this is where careful users get caught.

Four researchers varied only the formatting of a prompt (separators, capitalization, spacing) while keeping the meaning the same. Across 53 classification tasks and 320 formats, the accuracy spread for GPT-3.5 reached 56 points in the worst task, with a median spread of 6.4 points. For LLaMA-2-13B the spread reached 76 points. Making the model bigger, adding more examples, or using instruction-tuned versions did not eliminate the sensitivity¹.

A six-author team tested four formats (plain text, Markdown, YAML, JSON) on four OpenAI models. GPT-3.5 performance varied by up to 40% on a code-translation task depending on the template, and only 16% of its responses were identical between Markdown and JSON. Larger models were generally more robust, though one GPT-4 variant scored 76.2% on a coding benchmark with plain text and 21.95% with JSON, a gap the authors suggest came from the model writing reasoning text instead of code². This second paper is an unreviewed preprint and tested only OpenAI models².

Put these two bodies of evidence together and the lesson is not "use format X." It is that effects are real but model-specific and hard to predict, so the same prompt can behave differently after a model upgrade. That is why test-before-trust matters (Mistake 5).

Mistake 3: Leaving out what a new colleague would need

The strongest consistent advice comes from the model makers' own documentation, which is guidance rather than controlled research, but is notable for how much it agrees with itself. Anthropic's current guide tells users to think of the model as "a brilliant but new employee who lacks context on your norms and workflows," and offers a test: show the prompt to a colleague with minimal context, and "if they'd be confused, Claude will be too"⁹. It also recommends explaining why an instruction exists (its example turns "NEVER use ellipses" into an explanation that the text will be read aloud by a speech engine), stating what to do rather than what to avoid, and including three to five varied examples⁹. On shouting, it notes that newer models "might overtrigger on instructions that were needed for previous models," and suggests replacing "CRITICAL: You MUST use this tool when…" with plain wording⁹.

OpenAI's guide says the same in its own words: its models "benefit from more explicit instructions around how to accomplish tasks"¹⁰. Anthropic's engineering team describes the target as the "right altitude," between brittle, over-detailed rules and vague guidance that "falsely assumes shared context," and advises starting with a minimal prompt and adding examples where failures are observed⁸.

Note that this is the evidence grade of vendor documentation, not an independent trial. It is still the best available description of what these systems are tuned to expect.

Mistake 4: Stuffing the context window

More material is not always better. A peer-reviewed study from Stanford, UC Berkeley and Samaya AI researchers, "Lost in the Middle," gave models a question plus 10 to 30 retrieved documents and moved the document containing the answer around. Performance was highest when the answer sat at the beginning or end of the input and fell when it sat in the middle. GPT-3.5-Turbo's accuracy could drop by more than 20%, and in the worst cases with 20 or 30 documents it did worse than having no documents at all (56.1% closed-book)⁶. Those were 2023 models, and newer ones have improved on some tests, so treat the numbers as an illustration of a failure mode.

A 2025 technical report from Chroma, a company in the retrieval-database business and therefore with an interest in this topic, tested 18 models across three tasks (194,480 calls in total) and found that models "do not use their context uniformly" and that "their performance grows increasingly unreliable as input length grows." The authors stress their tasks were intentionally simple, and that real applications are harder⁷. Anthropic's engineering team now describes context as "a finite resource with diminishing marginal returns" and the degradation as "a performance gradient rather than a hard cliff"⁸.

The takeaway is to include what the task needs and trim the rest, with the key instructions and facts placed where the model is least likely to lose them.

Mistake 5: Never testing

Every study above had to measure its way to its answer, and none of the folklore survived unconditionally. Yet most prompts in daily use are never tested at all. OpenAI's guidance recommends pinning applications to specific model snapshots and "building tests and evaluation suites that measure prompt behavior so you can monitor performance as you iterate, or when you change and upgrade model versions"¹⁰. Given the format-sensitivity results¹², that advice is not bureaucracy. It is the only protection against a prompt that quietly stops working.

Limits and open questions

Most controlled prompting studies use multiple-choice exams, which are not representative of drafting emails, summarizing contracts or analyzing spreadsheets. Several results are from 2023 to 2025 models and may not carry over to current ones. The Wharton reports test a limited set of models and apply no correction for running many comparisons³⁵. The formatting studies cover classification and coding rather than open-ended writing¹². And for the context studies, nobody has a settled explanation of why long inputs degrade performance, which makes it hard to predict when a new model will fix it⁷.

What this means for you

Here are actions you can take this week.

1. Run the colleague test. Give your most-used prompt to a coworker unfamiliar with the task. Wherever they ask a question, add that answer to the prompt.
2. Replace one "don't" with a "do plus why." Rewrite a negative rule as a positive instruction and a one-sentence reason.
3. Drop the incantations, then check. Remove persona lines and "think step by step" from a prompt you use often, run both versions on ten real examples, and keep whichever scores better. If you use a reasoning model, expect little change.
4. Add two or three varied examples of the output you want, drawn from your real work.
5. Trim the context. Cut documents and history that the task does not need, and put the essential instruction and facts at the start or end.
6. Keep a small test set of 10 to 20 real cases and re-run it whenever your tool's model changes.

Sources

1. Sclar, M., Choi, Y., Tsvetkov, Y., & Suhr, A. (2024). Quantifying language models' sensitivity to spurious features in prompt design or: How I learned to start worrying about prompt formatting. ICLR 2024 (arXiv:2310.11324). arxiv.org/abs/2310.11324
2. He, J., Rungta, M., Koleczek, D., Sekhon, A., Wang, F. X., & Hasan, S. (2024). Does prompt formatting have any impact on LLM performance? arXiv:2411.10541 (preprint). arxiv.org/abs/2411.10541
3. Basil, S., Shapiro, I., Shapiro, D., Mollick, E., Mollick, L., & Meincke, L. (2025). Prompting Science Report 4: Playing pretend: Expert personas don't improve factual accuracy. Wharton Generative AI Labs, arXiv:2512.05858. arxiv.org/abs/2512.05858
4. Meincke, L., Mollick, E., Mollick, L., & Shapiro, D. (2025). Prompting Science Report 2: The decreasing value of chain of thought in prompting. Wharton Generative AI Labs, arXiv:2506.07142. arxiv.org/abs/2506.07142
5. Meincke, L., Mollick, E., Mollick, L., & Shapiro, D. (2025). Prompting Science Report 3: I'll pay you or I'll kill you, but will you care? Wharton Generative AI Labs, arXiv:2508.00614. arxiv.org/abs/2508.00614
6. Liu, N. F., Lin, K., Hewitt, J., Paranjape, A., Bevilacqua, M., Petroni, F., & Liang, P. (accepted 2023). Lost in the middle: How language models use long contexts. Transactions of the Association for Computational Linguistics (arXiv:2307.03172). arxiv.org/abs/2307.03172
7. Hong, K., Troynikov, A., & Huber, J. (2025, July 14). Context rot: How increasing input tokens impacts LLM performance. Chroma technical report. trychroma.com/research/context-rot
8. Anthropic Applied AI team. (2025, September 29). Effective context engineering for AI agents. Anthropic Engineering. anthropic.com/engineering/effective-context-engineering-for-ai-agents
9. Anthropic. (accessed 2026-11-08). Prompting best practices. Claude Platform documentation. platform.claude.com
10. OpenAI. (accessed 2026-11-08). Prompt engineering. OpenAI API documentation. developers.openai.com/api/docs/guides/prompt-engineering