Heidelberg AI News

Video · David Ondrej ·

More skills = worse agents (do this instead)

In short

The video argues that copying large viral skill repositories can overload an AI agent’s context, waste tokens, and reduce reliability. Instead, keep only a small set of trusted, task-specific skills, create custom skills from recurring workflows, and separate global skills from project-specific ones.

Key points

  • Delete roughly 80–90% of unused skills after reviewing session-usage data.
  • Treat skills as agent SOPs: repeatable workflows encoded for a specific task.
  • Prefer specialized project skills over vague global skills that load everywhere.
  • Use external skills only when they come from a trusted expert or organization.
  • Use a second frontier model to review substantial changes made by the first.
  • Create skills when a workflow is recurring, specialized, or likely to be repeated.
  • Share domain-specific skills through team repositories so expertise becomes reusable.

Written by an AI model from the video’s captions and description.

In this video, David Ondrej explains how he uses agent skills for development and why you should not be copying other people's skills but instead creating your own. He also goes into how to develop code using cloud agents inside of Cloudroom.

Watch on YouTube The opening of the article at youtube.com.

In this story

A tool or model opens in the Heidelberg AI Curriculum: what it does, what it costs, how to start. A repo opens on GitHub.

Related stories

All stories

VideoDavid Ondrej

The AI bubble is popping

The video argues that the AI market may be correcting, but the underlying technology is accelerating as frontier models improve and AI adoption expands. For AI builders, the advantage is speed: use leading models across development, security, and operations, and build products that benefit from agents rather than resist them.

VideoDavid Ondrej

My Agentic Engineering Workflow (after 15,700 sessions)

The video presents a fully agent-driven software workflow using frontier models, focused on rapid production shipping, cross-model code review, and isolated cloud execution. It argues that engineers should think like architects, delegate most coding to agents, and use human judgment mainly for task definition, risk assessment, and review.

VideoDavid Ondrej

Build your own open-source Grok Bot, here’s how

The video demonstrates building an open-source, self-hosted alternative to Grok Bot by reverse-engineering its UI and architecture with AI coding agents. It shows how multi-agent workflows, Rust, cloud sandboxes, and VPS deployment can accelerate development while keeping humans responsible for product and architecture decisions.

VideoDavid Ondrej

My $1,000 Fable 5.2 Workflow (just copy me)

The video explains a coding workflow that combines Fable 5.2 for difficult planning and reasoning, Opus 5.5 for faster execution, and Gemini 4 Carbon for independent reviews. It matters because Fable is highly capable but consumes limits quickly, so model selection and task-based routing reduce cost and latency.