Descriptions:
David Ondrej walks through his open-source agent skills repository — a collection of over 42 battle-tested skill files for AI coding agents including Claude Code, OpenAI Codex, Hermes, and Pi Agent — which went viral on GitHub and Twitter after he published it, attracting praise from the co-founder of Hermes Agent, the CTO of Boldmetrics, and the CTO of OpenSea.
The video focuses on eight of Ondrej’s favorite skills, starting with global agent guardrails: a hooks-based safety layer designed for use when running agents in “Yolo mode” (auto-approval of all commands). The skill intercepts dangerous shell patterns — recursive deletes, piping the internet into shell, rewriting remote Git history — before they execute. From there, Ondrej covers goal loop architecture, explaining how to translate vague instructions into verifiable end-state conditions that keep agents working autonomously until a specific, measurable outcome is reached. He demonstrates this live with Claude Code, showing a two-sentence plain-English prompt being converted into a structured goal loop with defined output formats and stopping conditions.
A throughline of the video is Ondrej’s argument that writing effective agent skills — instructions crafted for AI consumption rather than human reading — is becoming one of the most valuable engineering skills of 2026 and beyond. The full repository is available at github.com/davidandr/skills and covers orchestration patterns applicable across all major agentic frameworks.
📺 Source: David Ondrej · Published August 02, 2026
🏷️ Format: Hands On Build
