Agentic Coding

This video provides comprehensive guidance on configuring AI agents for efficient coding workflows. It covers the critical components that get sent to your AI with each request: system prompts, agents.md/claude.md files, skills, documentation, and your actual prompt. The key takeaway is keeping your agents.md file small and focused by extracting large sections into separate documentation files that you reference rather than include directly. This reduces token waste and prevents context pollution. The video demonstrates breaking down bloated agents.md files (like those from major companies containing thousands of tokens) into a lean structure with just a brief project description, package manager info, and links to specific conventions for API, web, testing, React, TypeScript, and delivery workflows. For monorepo structures, you can create nested agents.md files in different folders so the AI loads only relevant configuration when working in specific packages. Skills serve as on-demand documentation for specific use cases like authentication—loaded only when needed rather than always present. The speaker emphasizes starting with a minimal agents.md and adding conventions only when you notice the AI making mistakes, rather than preemptively overloading it with information.

Matt Pocock's skills repository provides a comprehensive set of agent patterns designed to address common failure modes when working with Claude Code and other AI coding agents. The skills are organized around four core problems: misalignment between developer intent and agent output (fixed with /grill-me and /grill-with-docs), excessive verbosity (addressed through building a shared domain language in CONTEXT.md), poor code quality (managed via red-green-refactor TDD loops and /diagnosing-bugs), and architectural decay (prevented through /improve-codebase-architecture scans). Installation is straightforward via Claude Code's plugin marketplace or as editable files in your project. The framework emphasizes small, composable, adaptable skills rooted in software engineering fundamentals like domain-driven design, test-driven development, and deliberate architecture care—enabling solo founders and teams to maintain code quality and developer control while leveraging AI acceleration.

This video demonstrates testing an "I have ADHD" skill for Claude Code that reformats AI responses to prioritize action over explanation. The skill applies 10 rules including leading with next actions, numbering multi-step tasks, suppressing tangents, and eliminating preambles. Across 18 real Claude Code sessions, responses with the skill were 36% shorter while maintaining correctness, with significant improvements in error reporting and debugging workflows. The skill reduced reading load from 1,809 to 1,164 words across test cases, showed clear next steps in 5 of 9 answers versus zero without it, and achieved a weighted benchmark score improvement from 4.05 to 4.47. Installation is simple—paste one sentence into your agent or install as a plugin, then activate with "/I have ADHD" command. The main limitations are that the model sometimes ignores rules, the skill doesn't prevent destructive commands, and overly short final messages can obscure reasoning. For solo developers using Claude Code daily, the skill notably improves debugging and error handling workflows while remaining free and MIT-licensed across multiple AI coding assistants.

APM is an open-source dependency manager for AI agents that addresses a core pain point for agentic coding workflows. It uses a single apm.yml manifest to declare and reproduce agent configuration across Claude, Copilot, Cursor, and other AI coding tools — similar to package.json or requirements.txt but for agent context, skills, plugins, and MCP servers. Key features include portable manifests with transitive dependency resolution, security scanning for compromised packages and Unicode hijacking, integrity hashing via lockfiles, and organizational policy enforcement through apm-policy.yml. For solo founders and small teams using Claude Code and Cursor daily, APM streamlines the setup of AGENTS.md files, skills, and prompts by making configurations shareable and reproducible across developers and machines without manual reconfiguration.