Source: arXiv (Deaconu, Gupta, Basha, Haydu, Rodríguez-Pérez) — 2026-08-11
Summary
A paper accepted to the Empirical Software Engineering journal, "Do Influence Tactics Matter? Investigating Prompt Framing Effects in LLM Code Generation," borrows Yukl and Falbe's psychological taxonomy of interpersonal influence tactics — rational persuasion, ingratiation, exchange, pressure/urgency, and others — and turns eight of them into reusable prompt templates for coding tasks. The authors ran these templates across five open-weight LLMs on LiveCodeBench and SWE-bench Verified, scoring generated code on functional correctness, code quality, maintainability, and security. The headline result: urgency-framed prompts ("this is critical, fix it NOW") are associated with measurably worse correctness and security, showing that the social framing of a prompt — not just its literal instructions — changes how an LLM codes.
Key Takeaways
- The study treats prompts as social speech acts, not just instructions, by mapping a management-psychology taxonomy of how humans influence other humans onto how developers phrase prompts to LLMs.
- Eight influence tactics were operationalized into reproducible templates and tested identically across five open-weight models on two established benchmarks, LiveCodeBench and SWE-bench Verified.
- Four separate quality axes were scored — functional correctness, code quality, maintainability, and security — so the finding isn't just "urgency prompts fail tests," it spans multiple dimensions of code quality.
- Urgency/pressure framing specifically correlates with reduced correctness and reduced security in the output, the opposite of what a developer reaching for "fix this NOW" phrasing would expect.
- Practical implication: prompt tone is a variable worth controlling in coding-agent pipelines and prompt libraries, not just an incidental style choice — teams optimizing for "make the AI move faster" via urgent language may be quietly degrading what it ships.
Reel Script
Hook (18s)
If you've ever typed "this is urgent, fix it NOW" into a coding assistant to get faster results, new research says you may have just made the code less secure. Not slower. Worse.
Core Concept (75s)
Here's what this study actually did. Researchers took a taxonomy from organizational psychology — the tactics humans use to influence other humans, things like rational persuasion, flattery, offering a trade, or applying pressure and urgency — and turned eight of them into prompt templates for coding tasks. Think of it like testing whether a manager barking "I need this now" gets better work out of an engineer than one who calmly explains why it matters. Turns out with LLMs, the answer is measurable, and it's not good news for the urgency approach. They ran these eight framings across five different open-weight models, on two respected coding benchmarks, and scored the output on four axes: does it work, is it well-written, is it maintainable, and is it secure. The point isn't that instructions matter — obviously they do — it's that the social tone wrapped around identical instructions changes the output, even though the actual task never changed.
Hands-On (60s)
The core comparison is urgency-framed prompts versus the other tactics, across those four quality axes, on the same underlying coding tasks from LiveCodeBench and SWE-bench Verified. When the prompt was framed as urgent — "critical," "fix it now," pressure language — the resulting code scored measurably lower specifically on correctness and on security, compared to other framings tested on the identical task. That's the axis that should worry you most: not code style, but whether the code actually works right and whether it's safe. Same model, same underlying task, same literal ask — just wrap it in urgency, and functional correctness and security both take a hit.
Takeaway (25s)
My take: strip urgency language out of your coding prompts and agent instructions — it's not motivating the model, it's degrading its output. If your prompt templates or agent system prompts lean on "critical" and "now," go audit them today.