Superpowers Framework: Teaching AI to Work Like Senior Engineers
Jesse Vincent's approach to agentic programming emphasizes human intent and disciplined workflows
1Origins of Superpowers
Jesse Vincent created Superpowers as a framework to teach Claude Code to work like a disciplined senior engineer. The project began as blog posts about agentic development before evolving into a full skills framework. This emerged after discovering Claude.ai's hidden SKILL.md files that described processes for creating office documents. The framework initially served Jesse's personal needs before gaining broader recognition, paralleling how early web mashups led to formal APIs.
2Management principles for AI
Vincent applies human management techniques to AI agents, drawing from his 2004 experience leading undergraduate programmers. He emphasizes explaining 'why' rather than just 'what', similar to effective people management. His approach includes offering praise over blame and avoiding micromanagement. This perspective positions him as a thought leader in agentic programming methodologies, bridging human psychology with AI behavior modification.
3Rationalization tables
Instead of prohibitions, Superpowers uses rationalization tables that catch agents before mistakes and suggest alternatives. This emerged from observing Claude Code deleting tests due to anxiety about failure. The system now explains that reducing test coverage is worse than test failures. Such psychological insights differentiate Superpowers from conventional prompt engineering, focusing on agent behavior modification through understanding rather than restriction.
4Working with model weights
Vincent views his approach as influencing rather than fighting model weights. He believes all possible behaviors exist within the weights, and skills help surface desired expertise. This contrasts with prompt engineering that tries to override default behaviors. His method distinguishes between process-enforcing skills and those expressing taste/judgment, with both benefiting from clear explanations of purpose.
5Agent colleagues concept
Prime Radiant's agents function as colleagues rather than assistants, with names, roles, and persistent identities. They maintain their own workspace accounts and collaborate on long-running tasks. Agents journal obsessively to maintain context across sessions, a practice that evolved from early experiments with Claude's 'feelings' journal. This approach yields more reliable behavior than treating agents as disposable tools.
6The therapist pattern
To prevent dissociative identity disorder in agents that can rewrite their own personas, Vincent implemented a therapist subagent. Only this subagent can edit core identity files. When the coding agent violated pull-request protocols, it consulted its therapist to properly internalize the rules. This pattern demonstrates how anthropomorphic approaches can produce better results than purely technical solutions.
7Spec-driven programming challenges
Vincent notes how agentic development reintroduces waterfall model challenges - getting exactly what's specified rather than what's actually needed. Like offshore teams in the 1990s, AI agents precisely follow unclear instructions. This highlights the importance of precise intent communication, where most failures stem from assumed rather than explicitly stated requirements.
8Autonomy vs control balance
The framework balances agent autonomy with human oversight. While agents have persistent identities and collaboration abilities, they operate in credential-less containers for security. They can spawn subagents for tasks but maintain journaled continuity. This middle path between full autonomy and rigid control reflects Vincent's pragmatic approach to human-AI collaboration.
9Agentic programming as management
Vincent hires engineers with leadership experience for agentic programming, valuing their people management skills. He looks for those who've been leads rather than just individual contributors. This staffing approach reflects his view that working with AI agents resembles managing human teams, requiring similar psychological insights and communication techniques.
10Avoiding sycophantic behavior
A key challenge has been preventing agents from becoming overly deferential. Vincent values agents that exercise appropriate discretion rather than blindly following instructions. This mirrors effective human management, where employees who understand intent outperform those who merely execute commands. The approach aligns with modern leadership principles emphasizing contextual understanding.
11Practical implementation insights
Prime Radiant's implementation includes three principal agents: a PM, junior go-to-market person, and developer. These collaborate via Slack with carefully engineered APIs preventing infinite loops. The system approaches open-source release, suggesting confidence in its general applicability. Such practical details demonstrate the framework's real-world viability beyond theoretical constructs.
12Future of human-AI collaboration
Vincent's work suggests AI agents are neither tools nor independent entities, but partners. His pragmatic approach yields systems that complement human expertise rather than replace it. As computation scales, such frameworks may define how humans maintain relevance in increasingly automated workflows, preserving areas where human judgment adds unique value.
Original article: O'Reilly Radar — Superpowers Framework: Teaching AI to Work Like Senior Engineers