From “Vibe Coding” to True Engineering: The Essence of Matt Pocock’s AI Agent Extension Tool “skills”

In 2026, the engineering landscape has been transformed by the widespread adoption of AI agents like GitHub Copilot and Claude Code. Giving instructions in natural language and receiving immediately functional code—so-called “Vibe Coding”—is now a daily part of development. However, in professional settings, a serious challenge has come to light: the “thinness of design density” and “lack of context” in AI-generated code.

The focus of this article, “skills”, is a masterpiece of prompt engineering built by Matt Pocock, a world-renowned authority on TypeScript, to optimize his own development process. It is more than just a convenience tool. It is an innovative framework that elevates AI from a “tool that merely waits for instructions” to a “partner that refines designs together.”

Tech Watch Perspective: The true value of this tool lies not in increasing the AI's generative capacity, but in forcing a workflow that builds a "Shared Language (Context)" between human and AI. The reason many developers fail when delegating tasks to AI is "ambiguity in specifications." The `/grill-me` command included in skills offers a reversal of logic: by having the AI "interrogate" the user, it fills design gaps before implementation begins.

1. Why Current AI Agents Need “Skills”

While AI agents possess incredible information processing capabilities, they suffer from two decisive bottlenecks:

  1. Contextual Disconnect: AI cannot automatically understand domain knowledge specific to a development site or architectural policies agreed upon within a team.
  2. Redundant Output: The more ambiguous the instructions, the more the AI outputs redundant explanations and generic code as a “safety measure,” wasting precious tokens and the developer’s focus.

Matt Pocock’s “skills” addresses these challenges by granting AI an “atomic, composable skill set.” This is essentially the act of installing “high-level protocols for performing specific tasks” into the AI’s “brain.”

2. Two Killer Skills That Dramatically Change the Developer Experience (DX)

① /grill-me: “Reverse Requirement Definition” to Expose Design Vulnerabilities

Normally, AI takes user instructions at face value. However, by executing /grill-me (or /grill-with-docs), the AI switches from “executor” mode to “reviewer” mode. Before starting implementation, the AI asks the user sharp questions such as, “How will this edge case be handled?” or “Is this data structure lacking scalability?” A few minutes of “sparring” before implementation prevents hours of debugging later.

② Shared Language: Reducing Cognitive Load via CONTEXT.md

This is a mechanism to share complex, project-specific concepts using short keywords. For example, by defining complex business logic in CONTEXT.md, the AI will perfectly grasp your intent the next time you simply say, “Apply that logic.” This strategy brings the concept of “Ubiquitous Language” from Domain-Driven Design (DDD) into the dialogue with AI, maximizing communication resolution.

3. The Decisive Difference from Existing Frameworks (GSD, BMAD, etc.)

Currently, frameworks like “GSD (Get Stuff Done)” are gaining attention for AI agent operations. These aim for full-process automation but often carry the risk of “the AI progressing too far on its own, becoming uncontrollable for the human.”

In contrast, “skills” remains committed to being a “toolbox that supports human decision-making.” The developer maintains leadership while calling upon necessary skills at the right timing. This design, predicated on “Human-in-the-loop,” is the primary reason “skills” is supported in professional environments.

4. Strategic Advice for Implementation

Best Practices for Adoption

  • Incremental Introduction: Don’t try to use every skill from the start. We recommend beginning with requirement clarification via /grill-me.
  • Dynamic Documentation: CONTEXT.md is not something you write once and forget. The key to success is establishing an operational flow, such as using an /update-docs command, where the AI itself updates the documentation as the project evolves.

Practical FAQ

  • Q: Does it depend on a specific model?
    • A: No. Because these are prompt-based abstracted skills, they are effective not only with Claude 3.5 Sonnet and GPT-4o but also with high-performance local LLMs.
  • Q: Can it be applied to existing large-scale projects?
    • A: Yes. In fact, the more complex the codebase, the greater the benefits of establishing a shared language.

5. Conclusion: Symbiosis with AI is Determined by the Quality of the “Question”

The quality required of future engineers is not the ability to memorize syntax. It is the ability to quickly build the “playing field” where the AI can produce high-precision output.

Matt Pocock’s “skills” is not just a collection of prompts; it provides a guide for “new craftsmanship” in the AI era. The phase of writing code based on “vibes” is over. Start educating your AI as a true partner today and step into the depths of design together. Beyond that lies a true engineering experience that can amplify an individual’s capabilities tenfold or a hundredfold.

This article is also available in Japanese.