What Happened
A new guide to "skills engineering" is making the rounds on Hacker News, and it points to a shift in how developers build capable AI agents. The core idea: instead of cramming everything into a system prompt, you package agent knowledge into modular SKILL.md files that load on demand.
The approach treats agent capabilities like software libraries. Each skill is a self-contained folder with metadata, instructions, and supporting resources. Agents load only what they need, when they need it - a pattern called "progressive disclosure" that keeps token usage efficient and context windows clean.
The ecosystem has grown fast. The community directory agentskill.sh now lists over 44,000 skills. Major platforms have adopted the standard: Claude Code (where Anthropic built native support), GitHub Copilot, Cursor, VS Code, and OpenAI Codex all support skill loading. Developer tools like SkillCheck (security scanning), OpenSkills (universal loader), and IntentKit (intent-driven building) have emerged around the format. Academic benchmarking via SkillsBench validates the approach across 86 tasks in 11 domains.
Why It Matters
If you've built AI agents, you know the pain of bloated system prompts. You stuff in instructions for every possible scenario, burn through tokens on context the agent rarely needs, and end up with brittle behavior when the prompt gets too long.
Skills engineering solves this by making agent knowledge modular and lazy-loaded. Your agent starts lightweight, then pulls in specific capabilities as the task demands. Need code review? Load the code-review skill. Need to write a blog post? Load the content-creation skill. The agent's context stays focused on what actually matters for the current task.
For teams, this means agent capabilities become version-controlled, reviewable, and shareable - just like code. You can distribute skills through shared directories, audit them before deployment, and build on each other's work instead of everyone maintaining their own massive system prompts.
Our Take
This isn't a new concept - it's a formalization of what good agent builders were already doing. But formalization matters. Having a standard format (SKILL.md with YAML metadata) means skills are portable across tools. A skill you write for Claude Code can work in Cursor or Copilot.
The real value is in the mental model shift. Prompt engineering asks "how do I tell the AI everything it needs to know?" Skills engineering asks "how do I organize knowledge so the AI can grab what it needs?" That's a fundamentally better question for complex agents.
If you're building agents that do more than one thing, stop writing monolithic prompts. Start thinking in skills. The tooling is mature enough in 2026 that there's no reason not to.