mattpocock/skills: A Practical Skillset for AI Agents

On July 15, 2026, Matt Pocock released a practical skillset for AI agents, titled “mattpocock/skills,” on GitHub. This project organizes essential agent skills for everyday engineering tasks and adopts a fundamentally different approach from “vibe coding.” While vibe coding emphasizes trial-and-error code generation, this skillset aims for controlled, reproducible workflows.

Matt Pocock, the creator, is a well-known engineer specializing in TypeScript’s type system and toolchain. He also runs a newsletter with approximately 60,000 subscribers, where updates about this project are shared.

Two Installation Methods

There are two main methods to install this skillset. The first involves using the skills.sh installer. Execute the following command in the terminal:

npx skills@latest add mattpocock/skills

Next, select the desired skills and target agent, then run /setup-matt-pocock-skills within the agent. This command interactively configures task management systems (GitHub, Linear, local files), triage labels, and document storage locations. Matt Pocock describes this as a “quick start that takes just 30 seconds.”

With this installation method, the skills are directly copied into the project’s repository, allowing users to freely edit and customize them. This approach is ideal for those who prefer managing skills on a project-by-project basis, similar to PHP’s composer or JavaScript’s npm.

The second method is to install the skillset as a native plugin for Claude Code. Run the following commands within Claude Code:

/plugin marketplace add mattpocock/skills /plugin install mattpocock-skills@mattpocock

Alternatively, execute these commands directly from the shell:

claude plugin marketplace add mattpocock/skills claude plugin install mattpocock-skills@mattpocock

With the plugin approach, the skills are introduced as read-only managed bundles. They are automatically updated whenever the author releases a new version, ensuring they remain up-to-date without user intervention. This method is ideal for users who prioritize maintenance-free operation over customization.

Matt Pocock positions the skills.sh method as suitable for users who want to “fork and enhance on their own,” while the plugin method is better for those who prefer to “subscribe and follow updates,” clearly delineating the philosophical differences between the two approaches.

The skills.sh installer is also compatible with other agents like Codex and supports the Agent-Skills standard harness. A native plugin for Codex is listed as a roadmap item for future development.

Why It Was Created

Matt Pocock developed this skillset to address common failure modes observed in coding agents like Claude Code and Codex.

The first failure mode is the issue of agents not behaving as expected. Quoting a passage from The Pragmatic Programmer, Pocock notes, “Nobody knows exactly what they want.”

A specific example is the misalignment between developers and agents. Developers may think they know what they want, but their intentions are not conveyed accurately to the agent, resulting in unintended outcomes. This problem persists in the AI era, manifesting as communication gaps between humans and agents.

As a solution, Pocock suggests a “grilling session,” where agents are repeatedly prompted with detailed questions to thoroughly extract the user’s intent and minimize misalignments. This skillset provides features to automate and streamline such interactive processes.

The original README also mentions second and third failure modes, but details are omitted in the initial release. Future updates will provide more information.

Differences from Existing Approaches

Existing frameworks for AI agents, such as GSD, BMAD, and Spec-Kit, attempt to support development by “owning” the entire process. However, Pocock identifies two key issues with this approach:

1. Users lose control over the process.
2. Resolving bugs within the process becomes difficult.

In contrast, this skillset is designed to be “small,” “adaptable,” and “composable.” It avoids locking users into specific processes and can be integrated with any model. Built on years of engineering experience, it encourages users to modify and personalize the skills as they see fit.

This approach shares a similar philosophy with the recently introduced agent-skills, which teaches development workflows to AI agents. While agent-skills serves as a general-purpose framework for educating agents, mattpocock/skills focuses on providing a specific set of practical skills. The two are complementary, allowing engineers to choose and combine them based on their workflows.

Conclusion

mattpocock/skills is a skillset designed for engineers who want to utilize AI agents as practical development tools. By offering ease of installation, freedom of customization, and continuous updates through plugins, it caters to diverse development styles.

Notably, it directly addresses the fundamental issue of misalignment between agents and users. This attempt to position agents as partners in the overall development process, rather than mere code completion tools, points to the future direction of AI-assisted development.

As agent-based development becomes more widespread, the importance of such practical skillsets is expected to grow.

Editorial Opinion

In the short term, we anticipate a rapid standardization of skillsets like mattpocock/skills across agents. With the expansion of Claude Code’s plugin marketplace, similar ecosystems may emerge for other agents like Codex and Cursor. Over the next 3–6 months, the “democratization of configuration” for AI agents is likely to progress, enabling individual developers to easily build optimized skillsets tailored to their needs.

In the long term, “skills” that control AI agent behavior could become the equivalent of traditional IDE plugins or package managers. In an era where agents autonomously generate and edit code, mechanisms to accurately convey human intent and ensure quality will be indispensable. The design philosophy of “small, composable skills” proposed by this project could become the standard architecture for AI development support tools within 1–3 years.

References

• “mattpocock /

  skills", by **mattpocock** — GitHub Trending, 2026-07-15 (ARR)

• Source URL: https://github.com/mattpocock/skills