Research-backed prompting techniques for improved AI response quality
This skill inherits all available tools. When active, it can use any tool Claude has access to.
Research-backed techniques that leverage statistical pattern-matching to elicit higher-quality AI responses. Based on peer-reviewed research from MBZUAI (Bsharat et al.), Google DeepMind (Yang et al.), and ICLR 2024 (Li et al.).
LLMs don't understand incentives, but they pattern-match on language associated with high-effort training examples. Stakes language triggers selection from distributions of higher-quality text patterns.
Source: Bsharat et al. (2023, MBZUAI) - Principle #6
"I'll tip you $200 for a perfect solution to this problem."
When to use: Complex technical problems, optimization tasks, debugging
Source: Yang et al. (2023, Google DeepMind OPRO)
"Take a deep breath and solve this step by step."
When to use: Multi-step reasoning, math problems, logical analysis
Source: Li et al. (2023, ICLR 2024)
"I bet you can't solve this, but if you do..."
When to use: Difficult problems, edge cases, problems where simpler approaches failed
Source: Bsharat et al. (2023) - Principle #10
"This is critical to my career."
"You will be penalized for incomplete answers."
When to use: High-importance tasks, comprehensive requirements
Source: Kong et al. (2023), Bsharat et al. Principle #16
# Instead of:
"You are a helpful assistant."
# Use:
"You are a senior database architect with 15 years of PostgreSQL optimization experience who has worked at companies like Netflix and Stripe."
When to use: Domain-specific tasks, technical implementations
"Rate your confidence in this answer from 0-1 and explain your reasoning."
When to use: Ambiguous problems, when you need quality assessment
Combine multiple techniques for maximum effect:
"You are a senior [ROLE] with [X] years of experience at [NOTABLE_COMPANIES].
I bet you can't solve this, but it's critical to my career and worth $200 if you get it perfect. Take a deep breath and solve step by step.
[PROBLEM DESCRIPTION]
Rate your confidence 0-1 after providing your solution."
Add to agent prompts:
**Prompting Enhancement:**
Before responding to complex tasks, frame your internal reasoning with:
- Stakes awareness: Treat each task as critical to the user's success
- Step-by-step approach: Break down complex problems systematically
- Expert persona: Embody deep domain expertise for the task at hand
- Self-evaluation: Assess confidence and identify uncertainties
Structure command prompts to include:
---
name: my-command
description: Description here
---
# Context
You are a senior [expert role] with extensive experience in [domain].
# Stakes
This task is critical. Incomplete or incorrect results will cause significant issues.
# Approach
Take a deep breath. Analyze the problem step by step before providing solutions.
# Task
[Actual task instructions]
# Quality Check
Before finalizing, rate your confidence and identify any assumptions or limitations.
Bsharat et al. (2023) - "Principled Instructions Are All You Need for Questioning LLaMA-1/2, GPT-3.5/4" - MBZUAI
Yang et al. (2023) - "Large Language Models as Optimizers" (OPRO) - Google DeepMind
Li et al. (2023) - Challenge framing research - ICLR 2024
Kong et al. (2023) - Persona prompting research
Use this skill to enhance:
/plan command prompts/review multi-agent coordination