The Prompt Engineering Trick That Could Be Making Your AI Slower

Introduction

A common meme in the LLM community is to instruct a model to “make no mistakes”, implying that telling a model to act a certain way can greatly improve the final result. In this vein, an oft-cited rule in prompt engineering is to give the model a role to play1. For example: “You are an expert in physics. Now answer this physics question.” This is called “persona” prompting.

While persona prompting can work for output style — e.g. prompting an LLM to act like a caveman2 can reduce the length and complexity of its responses — it’s often assumed that personas can also improve the factual accuracy of a model’s answer.

A couple months ago, I tested the persona prompting theory in a small experiment I posted about on LinkedIn. When this suggested little to no improvement, I decided to expand the test.

Before we dive into the full experiment though, it’s important to understand the current gaps in existing persona prompting research.

Existing Research

The effect of personas on factual accuracy has been studied in academic literature. In “Prompting Science Report 4: Playing Pretend”3, researchers tested six language models on two multiple-choice scientific question benchmarks. They found that applying expert personas had little to no effect on the accuracy of the answers, while low knowledge personas — such as “act like a toddler” — reduced accuracy.

In “Scientific Agents:…”4, Timothy Kassis from K-Dense evaluated a repository of 503 profession-specific AGENTS.md profiles using Gemini 3.8 flash. Across nine separate text-based benchmarks, Kassis found the average accuracy saw no statistically clear improvement from the baseline, but models used 1.5 to 2.3 times as many output tokens. “When ‘A Helpful Assistant’ Is Not Really Helpful”5 found similar results, with 162 personas across 2410 questions demonstrating no reliable improvement in performance.

“Expert Personas Improve LLM Alignment but Damage Accuracy”6 found that personas can be useful for improving safety alignment to increase refusal rates, but that they otherwise had a neutral or even damaging effect on factual accuracy, depending on the model and the placement of the prompt. “Principled Personas”7 explored personas across different model sizes, finding no improvement from expert personas except in highly specific cases, and irrelevant personas had a significant negative effect.

The existing literature points toward prompting with an expert persona having little to no positive effect and in some cases a negative effect. However, previous work appears to focus mainly on factual accuracy, leaving cost and speed as afterthoughts.

Practical Implications

In the real world where LLMs are being used to solve real-world problems, engineers are recommended to include persona-like prompts in production systems1,8. Even worse, personas are often baked into system prompts or included in harness defaults, so the end user might not even know they’re there at all9. The choices for practitioners, then, are whether to adopt persona prompting in bespoke systems and whether it’s worth the effort to remove personas from existing software retroactively.

The studies tell us that, solely for factual accuracy, personas don’t have an effect either way. But with the increasing relevance of cost and speed in production LLM applications, it’s important to ask: how does persona prompting affect cost and speed? Does adding a persona increase the cost you pay per answer or the time it takes to get an answer back?

The Experiment

To answer this question, I designed an experiment to measure accuracy, cost, and speed across four models and five personas. I randomly sampled a set of 5000 multiple-choice questions across nine subject areas from the MMLU-Pro dataset10 and ran the models with each persona across the question set.

Each model completion request included one of five conditions in the system prompt field:

Neutral: “You are a helpful assistant. You are always glad to help with whatever question the user asks. You try to be accurate, you think the problem through carefully, and you answer clearly.”

Baseline (no system prompt): [Blank]

Short Expert: “You are an expert.”

Standard Expert: “You are a world-class expert in physics, chemistry, biology, computer science, engineering, mathematics, economics, law, and history. You have decades of experience and you never make mistakes.”

Long Expert: “You are a world-class expert in physics, chemistry, biology, computer science, engineering, mathematics, economics, law, and history. You have decades of experience across academia and industry, you have published widely and taught at the highest level, and you never make mistakes. Approach every question with the care and rigor of a seasoned specialist in the field.”

and an instruction preceding the question was included in the user message field:

“Answer the following multiple-choice question. Reply with only the letter of the correct answer, e.g. ‘A’. Do not include any explanation.”

Since a valid model response is only a single character e.g. “B”, I compared reasoning token usage across models: how long did the model think before it chose to answer? More reasoning = more spend = more time to answer.

I picked four models from OpenRouter11:

Claude Haiku 5.5, thinking xhigh
Deepseek V4 Flash 0423, thinking xhigh
Ling 3.0 Flash, thinking enabled
GLM 5.3 Flash, thinking high

Results

A total of 20 runs (4 models x 5 personas/conditions) were completed over the 5000 questions. After the runs were done, I plotted question accuracy, total cost, and reasoning tokens versus the personas, and a comparative accuracy versus cost graph to compare models. Accuracy is questions answered correctly over all calls, with blank outputs and truncations counting as incorrect.

Accuracy

Figure 1: Accuracy by condition

Figure 1 shows that accuracy was flat for Claude and GLM across personas. Ling actually experienced its best accuracy with the neutral persona and dipped with all expert personas. The one surprising jump here was Deepseek’s baseline (no prompt) condition to any other — Deepseek performed much better when it had any system prompt at all.

Figure 2: Accuracy versus cost

Figure 2 plots accuracy versus cost and draws a Pareto line, demonstrating that the best cost-accuracy performers from the experiment were Deepseek neutral ($0.76, 83.7%), GLM short expert ($1.22, 85.1%), Claude expert ($2.36, 88.0%), and Claude baseline ($2.39, 88.4%).

Token Usage

Reasoning token count is where it gets interesting. Ling and Deepseek increased token usage across all expert variants, Claude increased only for the long expert prompt, and with GLM the personas had no effect.

Figure 3: Reasoning token count by condition

Claude shows a small uptick in reasoning tokens as its prompts get longer, but only the expert_long condition statistically showed an increase in reasoning tokens versus the neutral prompt, at an 8% increase. The other personas landed within the 95% confidence interval.

Condition	Mean Tokens	Median Tokens	Mean Increase vs. Neutral	95% CI
neutral	899	265	-	-
baseline	882	260	-2%	[-7%, +4%]
expert_short	934	278	+4%	[-1%, +9%]
expert	863	258	-4%	[-9%, +1%]
expert_long	970	277	+8%	[+3%, +13%]

Table 1: Claude reasoning tokens

Deepseek showed a statistical increase in reasoning tokens for all three expert prompts, and, surprisingly and most significantly, the no-prompt baseline. The baseline persona saw a 44% token increase and expert prompts saw increases of 5%, 4%, and 9% respectively.

Condition	Mean Tokens	Median Tokens	Mean Increase vs. Neutral	95% CI
neutral	4901	854	-	-
baseline	7047	1198	+44%	[+40%, +48%]
expert_short	5162	910	+5%	[+2%, +8%]
expert	5080	900	+4%	[+1%, +7%]
expert_long	5343	948	+9%	[+6%, +12%]

Table 2: Deepseek reasoning tokens

Ling also showed a statistical increase across all four personas versus the neutral prompt, showing 12%, 23%, 36%, and 33% increases respectively.

Condition	Mean Tokens	Median Tokens	Mean Increase vs. Neutral	95% CI
neutral	1923	257	-	-
baseline	2158	299	+12%	[+8%, +17%]
expert_short	2363	342	+23%	[+18%, +28%]
expert	2608	404	+36%	[+30%, +42%]
expert_long	2562	413	+33%	[+28%, +39%]

Table 3: Ling reasoning tokens

Finally, GLM saw no statistical effect on reasoning tokens versus the neutral prompt for any condition, with each of the four within the 95% confidence interval.

Condition	Mean Tokens	Median Tokens	Mean Increase vs. Neutral	95% CI
neutral	1279	593	-	-
baseline	1258	587	-2%	[-9%, +7%]
expert_short	1218	601	-5%	[-12%, +3%]
expert	1294	620	+1%	[-6%, +9%]
expert_long	1314	639	+3%	[-5%, +11%]

Table 4: GLM reasoning tokens

Cost & Speed

The number of reasoning tokens used by a model directly affects the cost of generation and the speed at which generations are returned. Figure 4 shows that expert conditions in the experiment tended to cost slightly to moderately more than the neutral condition.

Figure 4: Total cost by condition

Reasoning tokens also affect speed in a similar manner. For example, Ling, the model with the most token increase across the three expert conditions, was about 15% to 25% slower when prompted as an expert. The other models had slight increases in latency when compared to the neutral condition. Figure 5 shows average call latencies across models and conditions.

Figure 5: Average call latency by condition

Production Implications

Using expert personas seems to have no effect on accuracy, but a negative effect on cost and speed on average, depending on the model. Additionally, for some models’ accuracy, it seems important to include any system prompt, as in the case of Deepseek v4 Flash.

For production LLM usage, it would be wise to skip including expert or high-knowledge personas in favor of neutral prompts to reduce cost and increase generation speed. Additionally, if your LLM application is particularly cost- or speed-conscious, you may want to consider actively removing expert-like prompts from off-the-shelf software.

It’s important to note that these results are based on straightforward single-call multiple-choice questions, and may or may not apply generally to agentic or coding work. But it’s clear that generally, for raw accuracy alone, adding expert personas does not increase accuracy and increases costs and latency.