Agentic Coding

Matt Pocock's skills repository provides advanced agent engineering practices for Claude Code and Cursor users building production applications. The collection emphasizes alignment through grilling sessions (/grill-me and /grill-with-docs) to close communication gaps between developer intent and agent output, and introduces shared domain language via CONTEXT.md files to reduce verbosity and improve token efficiency. Key skills address common failure modes: /grill-with-docs combines requirement clarification with Architecture Decision Records, /tdd implements red-green-refactor loops for reliable code feedback, /diagnose provides structured debugging workflows, and /improve-codebase-architecture prevents complexity drift—all composable tools designed for real engineering rather than exploratory "vibe coding" and suitable for SaaS development workflows.

Vibe coding is a collaborative approach to software development where developers use AI tools in plain English conversation rather than writing code manually. The practical workflow involves four key tools: a conversational AI (like Claude or ChatGPT) for initial sketches, an AI-powered editor (like Cursor) for iterative development, GitHub for version control, and a deployment service (Cloudflare Pages, Vercel, or Netlify) for going live. Human judgment remains critical at three stages: crafting specific initial prompts, auditing the AI's output for subtle errors (especially security issues), and making scope and quality decisions. The article emphasizes that beginners don't need prior coding knowledge but benefit from being specific about requirements and patient with follow-up iterations. Real-world barriers like authentication bugs or database vulnerabilities require human review, making the process fundamentally collaborative rather than fully automated.

The website content provides a comprehensive CLI tool for managing agent skills across 51+ coding agents including Claude Code and Cursor. It documents how to install, create, and manage reusable instruction sets (skills) that extend AI coding agents' capabilities, such as generating release notes, creating PRs, and integrating with external tools. The skills are defined in SKILL.md files with YAML frontmatter and can be installed at project or global scope. The tool supports skill discovery, creation templates, and compatibility across different agents, directly addressing the user's interest in efficient workflows with Agent Skills and best practices for working with Claude Code and Cursor.

The article provides best practices for secure vibe coding workflows, which directly aligns with your interest in advanced AI assistant coding techniques. It emphasizes a five-layer security approach: keeping secrets out of code through environment variables and secrets managers, sandboxing code execution using microVM isolation (Kata Containers, Firecracker, gVisor) rather than standard containers, scoping database credentials to least-privilege access, isolating development and production environments with preview environments per pull request, and reviewing AI-generated output before deployment. The guide stresses that most vibe coding security failures occur at the deployment layer rather than the code layer, and recommends requesting security constraints explicitly in prompts ("Generate this with input validation, parameterised queries, least-privilege access, secure secret handling, and no hardcoded credentials") to improve the baseline of generated code. For solo founders using tools like Claude and Cursor, the article notes that deployment platforms like Northflank can apply these controls by default without requiring infrastructure expertise, allowing you to focus on rapid iteration while maintaining production-grade security.