The skill directory skills.sh, operated by Vercel, has indexed over 90,000 skill packs—equipping AI coding assistants with "methodologies" is becoming an essential requirement in the developer community.

What this is

We noticed that many teams using AI coding assistants encounter the same problem: AI writes code fast, but acts like an intern lacking engineering literacy—no tests, no design docs, and blind guessing when errors occur. Superpowers (an open-source AI coding skill framework) is here to fill this gap. It is essentially a set of software engineering methodologies that encapsulate the habits of senior engineers into individual "skill packs" (Skills, instruction files stored in SKILL.md format). Before starting work, the AI will first invoke a "brainstorming" skill to confirm requirements with you, then break down tasks. When writing code, it strictly follows TDD (Test-Driven Development: write a test and watch it fail, then write code to make it pass), and automatically triggers a code review afterward. To prevent excessively long instructions from blowing up the context, it employs a progressive disclosure strategy: normally, it only reads skill names and descriptions, loading specific instructions only when needed.

Industry view

The value of this approach is obvious: it elevates AI-assisted development from "generating code" to "doing engineering," effectively reducing technical debt. Currently, this skill ecosystem is quite prosperous. Vercel's skills.sh has indexed over 90,000 skills, with major players like Anthropic and Microsoft also contributing skill packs, supporting over 50 tools including Claude Code and Cursor. But the risks are worth our concern: the reliability of this framework depends entirely on the underlying LLM's instruction-following capability. If the model "disobeys," the best methodology becomes useless; meanwhile, for simple modification tasks, forcing the full workflow will instead increase token (LLM billing unit) consumption and wait times, dragging down efficiency. Some developers report that this system is currently better suited for medium-to-large requirements; using it for minor fixes is overkill.

Impact on regular people

For enterprise IT: the risk of introducing AI to write code shifts from "can the code run" to "is the process standardized." IT departments need to establish their own skill pack standards rather than letting employees install them at will. For individual careers: the core competency of programmers is shifting from "writing code yourself" to "managing AI's engineering discipline"—you must increasingly act like a tech lead. For the consumer market: ordinary users won't perceive its existence, but the probability of software co-developed by AI crashing due to low-level bugs will likely decrease significantly.