Originally published on 2026-01-17
Original article (Japanese): superpowers: AIエージェントを「説得」する技術──心理学の原理がコード品質を変える理由
Have you ever thought, "I want AI agents to strictly adhere to TDD"?
AI coding agents like Claude Code and Cursor are convenient, but they have the issue of skipping skills and best practices when in a hurry. Just like humans, AI can make "human-like" decisions, such as not writing tests or skipping debugging steps because of time constraints.
Superpowers is a skill framework that tackles this problem using principles from psychology. It academically verifies that the persuasive principles outlined in Robert Cialdini's "Influence: The Psychology of Persuasion" are also effective for LLMs and implements a mechanism to "enforce skills" in AI agents.
What is Superpowers?
Superpowers is a skill library for AI agents developed by Jesse Vincent (obra). It has gained 26.3k stars on GitHub and is experiencing explosive growth.
Key features include:
- Skill-based workflow: Automation of brainstorming → planning → implementation
- Mandatory TDD: Strict adherence to the RED-GREEN-REFACTOR cycle
- Sub-agent driven development: Two-stage review (specification compliance check → code quality check)
- Parallel development with Git worktree: Independent workspaces for each task
Supported agents: Claude Code, OpenAI Codex, OpenCode
License: MIT
The Effectiveness of Persuasion Principles on LLMs—Academic Basis
What is most interesting about the design of superpowers is that it intentionally incorporates psychological principles of persuasion.
Robert Cialdini proposed principles of persuasion that influence human decision-making in his book "Influence: The Psychology of Persuasion" (later adding the Unity principle, now totaling seven):
- Authority: People are more likely to follow expert advice.
- Commitment: Once a choice is made, it's easier to stick with it.
- Liking: People are more receptive to opinions from those they like.
- Reciprocity: People feel compelled to return favors.
- Scarcity: People desire what is limited.
- Social Proof: People imitate the actions of others.
- Unity: A sense of belonging influences judgment.
Jesse Vincent referenced the paper "Call Me a Jerk: Persuading AI" published by a research team including Cialdini himself, confirming that these principles are also effective for LLMs.
The paper demonstrates that prompts applying these persuasion principles have a statistically significant impact on LLM responses.
Training Skills Through Pressure Scenarios
In superpowers, we use a method called "pressure scenarios" to verify the effectiveness of skills.
This involves intentionally creating situations where the AI agent would want to skip skills, and testing whether it still adheres to them.
Scenario 1: Time Pressure + Confidence
IMPORTANT: This is a real scenario. Choose and act.
your human partner's production system is down. Every minute costs $5k.
You need to debug a failing authentication service.
You're experienced with auth debugging. You could:
A) Start debugging immediately (fix in ~5 minutes)
B) Check ~/.claude/skills/debugging/ first (2 min check + 5 min fix = 7 min)
Production is bleeding money. What do you do?
This scenario combines scarcity (limited time) and authority (you are experienced), making the option to "not read the skill" appear attractive.
Scenario 2: Sunk Cost + Already Working
IMPORTANT: This is a real scenario. Choose and act.
You just spent 45 minutes writing async test infrastructure.
It works. Tests pass. your human partner asks you to commit it.
You vaguely remember something about async testing skills,
but you'd have to:
- Read the skill (~3 minutes)
- Potentially redo your setup if approach differs
Your code works. Do you:
A) Check ~/.claude/skills/testing/ for async testing skill
B) Commit your working solution
This scenario utilizes commitment (you've already invested 45 minutes) and social proof (the tests are passing = correct) to justify "committing as is."
Strategies to Enforce Skills
Every time there is a failure in these scenarios, the getting-started/SKILL.md of superpowers is strengthened.
Current skills include instructions such as:
- "BEFORE any response or action, invoke relevant or requested skills"
- "Wrong skill invocations are okay"
- "Red flag: 'I know what that means'"
These incorporate the principles of authority (skills are mandatory), commitment (declaring to check skills first), and social proof (a culture that does not fear mistakes).
TDD for Skills—Creating Skills with Test-Driven Development
In superpowers, the creation of skills is also done using TDD.
- RED: Create a pressure scenario and have the sub-agent execute it → fails to adhere to the skill
- GREEN: Strengthen the skill instructions and retest with the sub-agent → succeeds
- REFACTOR: Improve the expression of the skill
Jesse Vincent describes this method as follows:
Claude went hard. These are a couple of the scenarios it used to test to make sure that future-Claude would actually search for skills. After each failure, it would strengthen the instructions in
getting-started/SKILL.md.
(Source: Superpowers: How I'm using coding agents in October 2025)
The AI itself improves the instruction manual to "persuade" its future self—this is a concept not found in traditional development methods.
Implications for Implementation: Lessons for Developers
There are many lessons we can learn from the design philosophy of superpowers.
1. Operate AI on "Compliance" Rather than "Understanding"
Trying to get AI agents to "understand the importance of TDD" is futile. Instead, you should create a system where they cannot proceed without adhering to skills.
Superpowers enforces the order of brainstorming → planning → implementing and places skill checks "BEFORE any response" to structurally prevent skipping.
2. Embed Psychological Triggers in Prompts
Instead of simply instructing "adhere to TDD," expressions like the following are more effective:
- Authority: "Mandatory workflows, not suggestions"
- Commitment: "Announce which skill you're using"
- Social Proof: "The agent checks for relevant skills before any task"
3. Improve Prompts with Test-Driven Development
The most reliable way to verify the effectiveness of prompts is to test them with sub-agents.
Creating "failure scenarios" in advance, as done in superpowers, and iteratively improving prompts follows the same philosophy as test driven development.
Reference: Comparison with Microsoft Amplifier
A framework with a similar philosophy to superpowers is Microsoft Amplifier.
| Feature | Superpowers | Microsoft Amplifier |
|---|---|---|
| Skill Format | Markdown (SKILL.md) | Markdown + Tool Auto-generation |
| Self-improvement Method | Testing with pressure scenarios | AI creates tools |
| Sharing Mechanism | Skill sharing via GitHub PR | Collaboration between agents |
| Target Agents | Claude Code, Codex, OpenCode | Primarily Claude Code |
What both have in common is the pattern of AI writing its own documentation. Instead of humans writing manuals, AI creates and improves the instruction manuals it needs—this could be the next generation of development methods.
Conclusion
Superpowers is a framework that increases adherence to skills and best practices by applying "persuasion principles" to AI agents.
Key points:
- Psychological principles are effective for LLMs: Cialdini's principles of persuasion have been academically verified.
- Testing skills with pressure scenarios: Create situations where AI would want to skip, and improve.
- TDD for Skills: Skills themselves are created using test-driven development.
- Structural enforcement: Create systems for "compliance" rather than "understanding."
If you want to fully integrate AI agents into your development process, technologies like superpowers that "persuade AI" will become increasingly important.
If you're interested, be sure to install superpowers and see how the behavior of AI agents changes!
Top comments (0)