GitHub's trending charts have recently been dominated by a single name: obra/superpowers, which rocketed to 270,000 stars within a week. According to Leiphone, it peaked at 1,422 new stars in a single day—roughly the daily traffic of two mid-sized open source communities.
It is not a model, nor an IDE plugin, but a skill framework + software development methodology for AI coding agents. The author, Jesse Vincent (obra), has packaged the engineering discipline the AI coding community has been developing over the past half year—ask about requirements first, write plans, test-drive, review before committing, verify before completion—into more than 20 composable skills.
How the Workflow Works
When an agent starts and sees that you want to build something, it doesn't immediately start writing code. Instead:
1. It steps back and asks what problem you're actually trying to solve 2. It refines requirements into a design document you can actually finish reading, confirmed with you in stages 3. After the design is signed off, it breaks work into 2–5 minute tasks, each with explicit file paths, code, and verification steps 4. It enters "subagent-driven development": a fresh subagent is dispatched for each engineering task, and completed work goes through a two-stage review (first checking spec compliance, then code quality), with critical issues blocking progress 5. Red-green TDD is enforced throughout: tests must fail first before implementation, and non-compliant code gets deleted and redone
Cross-Tool Portability
The most underrated feature is its cross-tool nature. It installs on 8 platforms: Claude Code's official plugin marketplace, Codex CLI, Codex App, Cursor, Gemini CLI, GitHub Copilot CLI, OpenCode, and even Kimi Code. Skills are written in Markdown + YAML frontmatter—readable by humans, parseable by machines—complete with dependency management and verification hooks. This effectively extracts "engineering experience" from model weights into a portable, testable, auditable asset.
Industry Implications
This is more worth pondering than the star count itself. Over the past year, the AI coding narrative has been "model power competition"—Claude vs. GPT, whose code is better. Superpowers' popularity twists the narrative in another direction: making agents reliable in engineering environments, rather than making models smarter. It even accelerates the commoditization of proprietary agent tools—skills being portable across platforms means neither Cursor nor Codex can lock in users; they can only compete on model quality and price.
This parallels Nvidia's recently disclosed AVO harness architecture as two sides of the same storyline: major vendors proving "the orchestration layer > the model itself" within their own systems, while the open source community spreads the same philosophy to all tools via skill frameworks. Top-down and bottom-up respectively, both point to the same judgment: AI coding's moat is migrating from model capability to the ability to write process into agents' muscle memory.
Three Signals to Watch
1. Whether the .agents directory convention can truly become a de facto cross-tool standard
2. Whether skill marketplaces develop paid distribution (what a "skill economy" business model looks like)
3. Whether Anthropic and OpenAI break their own capabilities into portable skills—if they do, it amounts to admitting this layer is public infrastructure, not any single company's private moat.