TechnicalDiscussion

AI Agent Skills Explained: What They Are and Which Ones to Use

Agent skills are reusable text instructions packaged in markdown files that teach AI agents how to perform specific tasks or follow processes, differing from prompts and GPTs by using progressive disclosure to avoid context bloat. The hosts discuss how skills like Grill with Docs, Wayfinder, and Code Review help structure AI workflows, emphasizing the importance of avoiding skill bloat while building reproducible processes across teams.

Summary

This episode explains that agent skills are markdown files containing natural language instructions for specific processes—not code libraries—that AI agents can selectively load based on context. Unlike traditional GPTs which load all instructions every time, skills use progressive disclosure, meaning agents first scan skill names and descriptions, then only load relevant details when needed, keeping context windows lean. The hosts clarify the distinction between skills (reusable process templates), agents.md files (core project rules and guidelines), and GPTs (separate agents with dedicated instructions and data).

The conversation emphasizes practical applications of skills: they create reproducibility in deterministic processes, enable team-wide standardization, and reduce token costs by avoiding redundant context loading. Matt Pocock's skills are highlighted as industry-leading examples, particularly Grill with Docs—a skill that asks 15-30 contextualized questions with recommendations during planning sessions, replacing simpler 2-3 question prompts. Grill with Docs also introduces 'with docs,' a terminology file that tracks project-specific vocabulary and acronyms, allowing the agent to understand company-specific language.

Other valuable skills discussed include Wayfinder (for identifying open-ended questions and challenges before implementation), To Spec (converting decisions into implementation specifications), To Tickets (breaking specs into actionable tickets), Implement (executing tickets with test boundaries), Code Review (checking for unnecessary complexity and regressions), and Handoff (preserving context for future sessions). The hosts warn against 'skill bloat'—downloading too many overlapping skills that increase context costs and create inconsistency. They recommend pruning skills monthly and creating custom skills based on team-specific workflows and processes. Computer Use skills are mentioned as more complex examples that can reference external scripts and assets. The conversation concludes that while skills will likely become standardized in team environments, implementation will vary by organization, and teams should adapt frameworks like Grill with Docs to Spec to Tickets to their specific needs rather than following them rigidly.

About this episode

Most of us spend far too much time teaching our AI coding agents the same lessons: ask questions before coding, write a specification, divide the work into tickets, test everything, and review the finished result. Agent skills let us package those instructions into reusable workflows that an agent can load whenever they’re needed.

Key Insights

  • Skills are simple markdown files—sometimes as few as seven lines—that guide AI agents through specific processes, not code libraries or complex implementations.
  • Progressive disclosure in skills means agents only load full skill details when relevant, unlike GPTs which load all instructions every time, making skills more token-efficient.
  • Grill with Docs transforms planning sessions from asking 2-3 questions to asking 15-30 contextualized questions, giving developers better project understanding and control through active decision participation.
  • The 'with docs' component of Grill with Docs creates a shared terminology file that tracks project-specific vocabulary and acronyms, allowing agents to understand company language across multiple sessions.
  • Skill bloat occurs when developers download too many overlapping skills from different sources, causing inconsistency and increasing token costs rather than improving performance.
  • Skills should be created from team processes (commits, deployments, testing, pull requests) that are repeatable and deterministic, translating operational patterns into reusable instructions.
  • The recommended workflow progression—Wayfinder → Grill with Docs → To Spec → To Tickets → Implement → Code Review—is a framework that teams may adapt rather than follow one-for-one based on their specific needs.
  • Computer Use skills represent complex skill types that reference external applications and scripts, requiring careful management to avoid overloading the agent with unnecessary capabilities.

Topics

Agent skills definition and structureSkills vs. GPTs vs. agents.md filesProgressive disclosure and context managementGrill with Docs skill and terminology filesSkill bloat and pruning strategiesTeam standardization of workflowsMatt Pocock's skill frameworkCustom skill creation based on processes

Transcript

Most of us spend a ridiculous amount of time teaching our AI coding agents the same lessons over and over. Ask questions before coding, write a spec, break the work into tickets, test it, and review it properly. Agent skills let us package those instructions into reusable workflows that an agent can load whenever it needs them. But not every skill is useful, and installing a pile of someone else's opinions can just straight up make your agent worse, not better. So today we're breaking down what agent skills actually are, how they differ from prompts, rules, and MCP servers, excuse me, and the handful of skills that I think are genuinely worth trying. And this is actually a Mike…

Full transcript available for MurmurCast members

Sign Up to Access

More from HTML All The Things - Web Development, AI, and Developer Careers

Get AI summaries like this delivered to your inbox daily

Get AI summaries delivered to your inbox

MurmurCast summarizes your YouTube channels, podcasts, and newsletters into one daily email digest.