Agentic Coding

Agent Skills is a curated library of 176 reusable SKILL.md workflows organized across nine categories (Codex, media, UI, illustration, game combat, game development, web design, and workflow) designed specifically for AI coding agents like Claude Code and Cursor. Each skill contains procedural steps, defaults, guardrails, and acceptance checks formatted as plain Markdown files that agents can load and follow. The library emphasizes prompt reusability, reference-driven design over descriptive text, and portable workflows that work across different projects without project-specific context. Key capabilities include video-to-prompt conversion, HTML-to-interaction-prompt workflows, full-page capture techniques, design systems (14 SVG illustration styles, combat VFX, Three.js game systems), web design patterns, and operational procedures for iterating with agents on real projects using scoring rubrics and progress screenshots.

Cursor and Claude Code have become primary coding tools for developers, with AI agents now handling most code writing tasks. The author describes shifting from writing code directly to directing AI agents, with Cursor discovering and fixing over 20 threading race conditions in minutes—issues that would be nearly impossible for humans to catch manually. This transformation highlights how AI coding agents are better equipped to handle complex reasoning tasks like multithreading and can be integrated into pull request workflows for quality assurance, fundamentally changing the developer's role from code author to software architect directing AI agents.

jCodeMunch is a token-efficient MCP server that cuts AI code exploration costs by 86-99% (averaging 96.5% reduction, 28.3x fewer tokens) by retrieving exact code symbols via tree-sitter AST parsing instead of loading entire files. It works with Claude Code, Cursor, Windsurf, Continue, and other MCP clients, enabling targeted edits on single functions to see ~95% token savings and offering structural queries (find_importers, get_blast_radius, get_call_hierarchy, find_dead_code) that native grep tools cannot answer. Installation is one command (uv tool install jcodemunch-mcp && jcodemunch-mcp init), and it includes an AGENT_HOOKS.md guide for configuring agents to prefer its tools and a CLAUDE.md prompt policy that stops agents from reverting to brute-reading entire files—critical for real efficiency gains in agentic workflows.