# superpowers-skills-coding-agents-vincent-2025-10-09

## Veille

AI Agent Skills System - Superpowers/Skills - Continuous Learning - Fsck Blog

## Titre Article

Superpowers: How I'm using coding agents in October 2025

## Date

2025-10-09

## URL

https://blog.fsck.com/2025/10/09/superpowers/

## Keywords

AI Skills System, Coding Agents Superpowers, Claude, Test-Driven Development, Git Worktrees, Subagents, Agent Learning, Skill Sharing, Self-Optimization, AI Workflow

## Authors

Jesse Vincent

## Ton

**Profile:** Conversational-Professional | First-person narrative | Educational-Opinion | Intermediate-Expert

Vincent continues his methodological series with the same practitioner-experimenter tone. The programmatic quote ("Skills are what give your agents Superpowers") reveals controlled enthusiasm for Skills systems. Documented iterative approach (Brainstorm → Plan → Implement, RED/GREEN TDD) with transparency about successes/failures. Assumes advanced development expertise (subagents, git worktrees automation). Consistent stylistic continuity with a personal blog sharing the evolution of field practices.

## Pense-betes

- Key quote: "Skills are what give your agents Superpowers"
- Skills system based on markdown documents
- Workflow: Brainstorm → Plan → Implement
- Automatic creation of git worktrees
- RED/GREEN TDD (Test-Driven Development) approach
- Dispatching tasks to subagents
- Code review integrated into the workflow
- Pressure testing skills via simulated scenarios
- Extracting skills from technical books
- Using persuasion principles to improve agent reliability
- Future developments: skill sharing, memory management, capability expansion
- Skills can be read, learned, implemented, tested, and shared

## RésuméDe400mots

Jesse Vincent presents a major evolution in his methodology for using AI coding agents with the concept of "Superpowers" - a structured skills system that allows AI agents to learn, improve, and continuously self-optimize. This approach radically transforms the way developers can collaborate with artificial intelligence.

The core of the system rests on skills formalized as markdown documents. These skills function as learning modules that AI agents can read, understand, and implement. Unlike simple prompts, these skills are persistent, testable, shareable, and constantly improvable. They literally give agents "superpowers," enabling them to execute complex tasks with specialized expertise.

The development workflow is structured around a three-phase methodology: Brainstorm (creative thinking and solution exploration), Plan (detailed planning and architecture), and Implement (concrete implementation). This linear process ensures that each step receives appropriate attention before moving to the next, avoiding rushed or poorly designed solutions.

Automation plays a central role in the system. Vincent integrated automatic creation of git worktrees, allowing the agent to work in isolated environments for each task. This isolation facilitates experimentation and minimizes the risk of conflicts or regressions in the main code.

The RED/GREEN TDD (Test-Driven Development) approach is deeply embedded in the workflow. The agent starts by writing failing tests (RED), then implements the minimal code needed to pass them (GREEN), thereby creating a robust test suite that guarantees the quality and reliability of the code produced.

A particularly interesting innovation is the dispatching system to subagents. The main agent can delegate specific tasks to specialized subagents, creating a collaborative architecture in which different AI instances work on different aspects of a project. This approach mirrors human development teams, with their specializations and collaborations.

Vincent has also developed techniques to test and improve the skills themselves. He uses "pressure testing" - subjecting skills to extreme simulated scenarios to identify their limits and weaknesses. He also explores extracting skills from technical books, transforming the codified wisdom of professional literature into executable instructions for AI agents.

Applying principles of psychological persuasion to prompts represents an innovative approach to improving agent reliability. By understanding how to formulate instructions to maximize their cognitive impact on the AI, Vincent optimizes success rates and result consistency.

Future prospects include mechanisms for sharing skills between developers, more sophisticated memory management enabling agents to maintain context over long periods, and the continuous expansion of agent capabilities into new areas of expertise.

In conclusion, the Superpowers system represents a paradigm shift in the use of AI coding agents, moving from one-off tools to continuous-learning partners equipped with specialized, evolving skills.

## GrapheDeConnaissance

- Jesse Vincent —a_créé→ Superpowers (TECHNOLOGIE, 0.98)
- Superpowers —utilise→ skills (CONCEPT, 0.97)
- skills —est_basé_sur→ fichiers SKILL.md (TECHNOLOGIE, 0.95)
- Anthropic —publie→ système de plugins Claude Code (TECHNOLOGIE, 0.96)
- Superpowers —utilise→ workflow Brainstorm-Plan-Implement (METHODOLOGIE, 0.95)
- Superpowers —utilise→ RED/GREEN TDD (METHODOLOGIE, 0.95)
- Superpowers —permet→ git worktrees (CONCEPT, 0.92)
- Jesse Vincent —utilise→ principes de persuasion Cialdini (CONCEPT, 0.9)
- principes de persuasion Cialdini —améliore→ fiabilité des agents IA (CONCEPT, 0.88)
- Superpowers —permet→ dispatching vers sous-agents (METHODOLOGIE, 0.93)
- Microsoft Amplifier —est_basé_sur→ auto-amélioration par fichiers markdown (METHODOLOGIE, 0.85)
- Jesse Vincent —collabore_avec→ Sam Schillace (PERSONNE, 0.75)
- Dan Shapiro —publie→ étude sur Cialdini et LLMs (DOCUMENT, 0.92)
- skills —s_inspire_de→ livres techniques (CONCEPT, 0.88)

---
Canonical: https://www.thekb.eu/en/fiches/superpowers-skills-coding-agents-vincent-2025-10-09/