Agentic Coding

These skills address core challenges in AI-assisted development. Key practices include using /grill-me and /grill-with-docs to establish alignment with the agent before coding, which involves asking detailed questions about requirements to bridge the communication gap. Building a shared language through a GLOSSARY.md document reduces verbosity and helps the agent use tokens more efficiently by using project-specific terminology consistently. For code quality, implement feedback loops using /tdd for red-green-refactor cycles and /diagnosing-bugs for disciplined debugging. To prevent architectural drift, run /improve-codebase-architecture regularly to identify deepening opportunities. The workflow emphasizes small, composable skills based on engineering fundamentals rather than process frameworks that remove developer control, with support for Claude Code, Cursor, and other agents through either managed plugin installation or editable local files.

This O'Reilly live event covers context engineering strategies specifically designed for agentic coding with Claude Code. Instructor Chelsea Troy teaches how LLMs produce more accurate answers through judicious context management, including practical techniques like using /clear, /compact, /exit, and /resume commands, @-files for narrowing context, and memory commands. The course addresses code characteristics that improve LLM tool effectiveness, how to summarize code bases with context documents, and advanced features like plan mode and think mode for structuring AI-assisted development workflows. Hands-on exercises include initializing Claude or Cursor on code bases, constructing plans for code changes, and optimizing the Claude file for better agent collaboration.

Claire's Codex automation workflow demonstrates an advanced agentic coding pattern: a meta-agent that periodically analyzes your GitHub pull request history to identify reusable skills, spawns sub-agents to validate each skill against your repository's main branch, and then adds proven skills to your agent's library. The approach runs on a recurring schedule (like weekly), with the main agent creating dedicated sub-agents for each proposed skill, each given explicit validation goals such as passing tests or successful builds. This self-improving loop reduces manual skill curation by automating the identification and testing phases, though it requires clear success metrics and occasional human judgment when validation fails or PR data becomes stale.

The GitHub page lists 284 repositories related to cursor-skills, many directly relevant to the user's interests in agentic coding with Claude Code and Cursor. Key resources include curated collections of 1000+ agent skills compatible with Claude Code and Cursor; AGENTS.md rules and frameworks inspired by clean code principles for AI coding agents; prompt templates for system prompts, tool prompts, agent delegation, and multi-agent coordination informed by Claude Code study; and SkillDock, a skill manager desktop app for organizing, editing, syncing, and updating Skills and MCP servers across Claude Code, Cursor, and other AI coding tools. Additional specialized collections cover 150+ software development skills, 380+ Claude Code skills and plugins with 30+ agents and 70+ custom commands, and lifecycle toolkits for mining repeated workflows and personalizing agent skills—directly addressing PRD workflow optimization and context management techniques for efficient AI-assisted development.