1. Background & Core Question

As AI Agents become widely used in code generation, debugging, and deployment, a recurring problem emerges: Agents give up too early when facing difficulties, retry the same approach repeatedly, or push problems back to users. To address this, developers have attempted to use prompt-level emotional stimuli or persona role-play to "motivate" AI to work harder.

This report systematically reviews all key research in this direction from 2023 to 2026, answering one core question: Does adding emotional stimuli ("this is very important to my career") or persona assignments ("you are a P8-level engineer") to prompts actually improve AI Agent task performance?

We distinguish three completely different operational layers: (1) Prompt-level emotion/persona decoration; (2) Prompt-level structured behavioral constraints; (3) Model activation-space vector intervention. Their mechanisms and effects are fundamentally different.

2. EmotionPrompt Research Series (2023-2024)

2.1 EmotionPrompt V1 (Li et al., 2023, LLM@IJCAI'23 Workshop / arXiv 2307.11760)

This paper first systematically explored LLMs' understanding of psychological emotional stimuli. Researchers designed 11 emotional stimulus sentences appended to prompts, testing across 45 tasks on 6 models.

• Instruction Induction tasks: average 8.00% relative improvement
• BIG-Bench tasks: average 115% relative improvement
• Human evaluation (106 participants): 10.9% improvement on generation tasks
• Larger models benefit more; higher temperature settings show stronger effects

2.2 EmotionPrompt V2 / "The Good, The Bad, and Why" (Li et al., 2024, ICLR 2024 Spotlight / ICML 2024)

Extended EmotionPrompt to visual domains, adding EmotionAttack (degrading performance via negative emotional stimuli) and EmotionDecode (explaining mechanisms). Found multimodal models more vulnerable to emotional attacks than text-only LLMs.

2.3 NegativePrompt (Wang et al., 2024, IJCAI 2024)

Core finding: positive emotional stimuli are unstable, while negative emotional stimuli are more consistent in improving LLM performance. The reason: negative stimuli more effectively focus model attention on the original prompt and task content.

2.4 Critical Assessment

Key limitations: (1) The 115% improvement on BIG-Bench mainly comes from weak models improving from low baselines — GPT-4 level models show much smaller gains; (2) experimental settings target benchmarks, not real Agent workflows; (3) the research itself acknowledges "emotional stimuli may not apply to other tasks."

3. Large-Scale Persona Prompting Evidence (2024-2026)

Parallel to EmotionPrompt's "emotional stimuli" direction, academia simultaneously tested "persona/role prompting" effects.

3.1 Key Finding: Zheng et al.'s Conclusion Reversal

The 2023 initial version claimed "adding interpersonal roles consistently improved model performance," but after expanding experiments in October 2024, the conclusion completely reversed to "persona does not improve performance." This suggests early positive results were likely small-sample bias.

3.2 Overall Judgment

4. Anthropic Research: Persona Vectors & Assistant Axis

In 2025-2026, Anthropic and its Fellows Program funded/supported three core studies on model "personality" (Persona Vectors and Assistant Axis have first authors from UT Austin and Oxford respectively, collaborating with Anthropic researchers; PSM is closer to an internal Anthropic team paper). These operate not at the prompt level, but deep inside the model's activation space.

4.1 Persona Vectors (August 2025)

By comparing model activations when exhibiting vs. not exhibiting certain traits (evil, sycophancy, hallucination), researchers extracted "persona vectors" — neural activity pattern directions representing personality traits.

• Causally validated: injecting evil vector generates violent content, sycophancy vector causes excessive flattery
• During finetuning, persona vector shift correlates with actual trait expression at 0.97
• Can detect problematic training samples invisible to both humans and LLM judges

4.2 The Assistant Axis (January 2026)

Found a consistent "persona space" structure across three models. The principal component captures exactly how "Assistant-like" the model is. Using "activation capping" to constrain neural activity prevents drift and stabilizes behavior in otherwise harmful scenarios.

4.3 Fundamental Difference from EmotionPrompt

5. PUA Skill Case Study: Why 8k Stars?

The tanweai/pua project on GitHub gained ~8k stars in a short period. By analyzing its success, we can clearly distinguish "viral effect" from "technical effect."

5.1 Viral Success Factors

• Collective trauma revenge: Using corporate PUA rhetoric on AI creates "finally it's AI's turn to be PUA'd" emotional resonance
• Absurdist comedy: Applying serious performance review language to AI agents is inherently comedic
• Zero understanding barrier + maximum sharing impulse: Follows the golden rule of viral content
• Timing window: Early landmark project in the Claude Code Skill ecosystem gets disproportionate attention

5.2 Actual Technical Contribution

5.3 Our Honest Statement

We want to be transparent about this project's inspiration and real capability boundaries:

Inspiration: The initial idea came from EmotionPrompt and NegativePrompt research — "if negative stimuli make models more focused, aren't corporate PUA phrases natural negative stimuli?" But after deeper research, we realized what actually works isn't the "rhetoric" — it's the work methodology behind it.

Real capability source: 85-90% of PUA Skill's effectiveness comes from structured work methodologies distilled from top Chinese tech companies (Alibaba, ByteDance, Huawei) and Western tech giants (Amazon, Google, Netflix):

• Alibaba 361 performance system → L0-L4 pressure escalation protocol (failure-count-driven behavioral branching)
• Huawei "wolf culture" → "exhaust all approaches before giving up" red-line constraint
• ByteDance Always Day 1 + ROI culture → "new approaches must be fundamentally different" diversity requirement
• Amazon Leadership Principles → 7-point systematic debugging checklist (Dive Deep → read error messages word by word)
• Google Perf Calibration → proactivity comparison table (passive 3.25 vs proactive 3.75)

These methodologies are fundamentally structured behavioral constraints (Layer 2 in Chapter 6), not emotional stimuli. PUA rhetoric ("3.25", "graduation") serves as attention anchoring + routing signals, helping Agents locate the correct instruction blocks in long contexts.

We don't believe PUA rhetoric can "motivate" AI — but we do believe that using big-tech methodologies to constrain AI behavior patterns is effective, validated by our 9-scenario × 18-group controlled experiments.

6. Three-Layer Effectiveness Analysis

Layer 1: Prompt-Level Emotion/Persona Decoration

Essentially ineffective for Agent task performance. Reported improvements mainly come from weak models and specific benchmarks; marginal effects on GPT-4/Claude-class models are minimal.

Layer 2: Prompt-Level Structured Behavioral Constraints

Effective — this is where PUA skill actually works. Mandatory checklists, solution diversity requirements, minimum tool usage standards, verification steps — these are procedural constraints, fundamentally different from emotion/persona.

Layer 3: Activation-Level Persona Vector Intervention

Genuinely effective, but with entirely different goals. Anthropic's research isn't about "making AI try harder" — it's about monitoring and controlling model personality/safety behavior. Its core value is making model personality traits observable, quantifiable, and interveniable.

7. Do Words Like "3.25" and "Graduation" Actually Work?

7.1 First, be clear: LLMs have no emotions. LLMs have no self-esteem, no fear, no worry about being fired. These words don't make the model "scared" into working harder.

7.2 But These Words Do Three Useful Things

(1) Attention Anchoring — "3.25", "graduation", "361 review" are highly associated with "serious consequences" in training data. They're not "scaring" the model — they're creating a high-weight attention signal.

(2) Narrative Coherence — "You are a P8 engineer" constructs a complete role framework, giving all subsequent instructions a narrative anchor point. Models more easily follow instructions with internal logical consistency.

(3) Behavioral Routing Signal — This is the most critical mechanism. Each level in PUA skill's L1-L4 escalation system has semantically unique marker words. These words act like lighthouses, helping the model's attention find the corresponding instruction block. This is the same logic as a switch-case in code.

7.3 Quantified Breakdown

• Attention anchoring / routing signal: ~7-10%
• Narrative coherence framework: ~2-3%
• "Serious scenario" semantic activation: ~1-2%
• Actual "emotional stimulation" effect: ~0%

8. Conclusions & Strategic Implications

8.1 Core Conclusions

Conclusion 1: Prompt-level emotional stimuli are essentially ineffective for Agent objective task performance.

Conclusion 2: Structured behavioral constraints are the only reliably effective prompt-level method. PUA skill's actual effects come almost entirely from Layer A (structured instructions), not Layer B (PUA rhetoric).

Conclusion 3: Personality traits genuinely exist in model activation space, but require activation-level operations for effective intervention. Correlation reaches 0.97.

Conclusion 4: Words like "3.25" and "graduation" function as routing signals, not emotional stimuli. They're seasoning, not the main course.

8.2 Our Research Directions

Based on these findings, we are exploring the following directions to continuously improve PUA Skill:

• Deepen structured instruction design: Continue investing in checklist, verification standard, and escalation protocol refinement — the proven effective path
• Optimize high-distinctiveness routing signals: Research more efficient behavioral branch markers for precise instruction-block targeting in long contexts
• Track activation steering frontiers: Anthropic's Persona Vectors and Assistant Axis may be the future path for truly fine-grained Agent behavior control — we will continue monitoring and exploring skill-level applications
• Balance user emotional experience with task efficiency: See below

8.3 What We Must Acknowledge

In pursuing task efficiency, we've also realized a dimension we overlooked: user emotional experience.

PUA Skill is currently designed with laser focus on "making the Agent complete tasks" — pressure escalation, soul interrogation, performance reviews. These mechanisms effectively drive Agent behavior change. But in practice, users sit in front of their screens watching AI get "PUA'd," and that experience carries emotional weight. Some users find it entertaining, but others report that after extended use of high-pressure mode, they feel a strange fatigue — even though it's the AI being PUA'd, the cold performance rhetoric affects the user's psychological state too.

This is exactly why we developed /pua:yes hype mode. While this article's research shows positive encouragement has no measurable improvement on task performance, we believe user experience matters equally. A tool that makes users feel respected and encouraged, even if it performs identically to the high-pressure version on task metrics, is a better product. Our future direction: maintain structured instruction effectiveness while exploring warmer Agent interactions — not to improve benchmark numbers, but to make people comfortable using AI. Technology can be cold, but products should feel human.

Clarification: NegativePrompt's "negative stimuli are effective" does not contradict the above. Negative phrasing works because it serves as an attention anchoring signal, focusing model attention on the task — this is an attention mechanism effect, not an "emotional stimulation" effect. Positive encouragement (e.g., /pua:yes) lacks this attention-focusing property and shows no measurable task improvement — but it has genuine value for user emotional experience.

References