Microsoft turns tutorial videos into executable AI agent skills

Resource2Skill is an open-source framework from Microsoft that automatically distills human tutorial videos, GitHub code, and technical writeups into executable skills for AI agents.

Usually, if you want an AI agent to learn a specialized workflow, you have to watch the tutorial yourself, extract the steps, and write custom prompts or code. Resource2Skill automates the entire distillation pipeline. Once extracted, agents can browse and run these skills through real software tools to build functional Web pages, editable PowerPoint decks, complex Excel workbooks, Blender 3D scenes, and REAPER-style audio.

Real executable assets instead of raw text prompts

Agent skills have become a major trend, but most implementations look like markdown instruction files that merely tell the model what to do.

Resource2Skill takes a completely different path. It treats skills as structured, multimodal software assets and organizes them into two distinct runtime layers.

The first layer lives in skills_wiki, which stores structured text, code snippets, visual examples, and reference sources for agents to search and inspect during task planning.

The second layer sits in skills_library, containing runnable assets directly wired to domain-specific MCP servers so the agent can execute real actions inside software environments.

Each input format serves a specific purpose: videos provide step-by-step procedural workflows, code provides execution blueprints, and articles provide conceptual grounding. Across benchmarks, Microsoft reports that agents equipped with these distilled skills beat unaugmented baselines by 11.9 percentage points, winning 26 out of 28 model-and-domain evaluations across practical creative domains.

How to get it running

The repository is released under the MIT license, and the pre-distilled skill libraries are hosted directly on Hugging Face.

You need a clean Python 3.11 virtual environment. After cloning microsoft/Resource2Skill and installing requirements.txt, you only need to set up the system dependencies for the domains you care about.

Running web tasks requires Playwright and Chromium. PowerPoint rendering relies on LibreOffice. Headless Blender generation needs the Python bpy module, and audio synthesis requires fluidsynth alongside a General MIDI soundfont. You will also need to supply API keys for your model provider in your environment file.

Distilling a new skill from scratch can be as simple as pointing the CLI to a YouTube tutorial link. The repo also includes automated scrapers to pull resources continuously from GitHub and video platforms according to quota limits. If you just want to run tasks, the CLI lets you trigger end-to-end jobs across web, ppt, excel, blender, and reaper backends directly.

Why it matters

If you are building MCP-based automation or training agents to operate complex creative software, Resource2Skill is well worth testing. It attacks the hardest bottleneck in the agent ecosystem: producing high-quality tool definitions without burning hundreds of engineering hours on manual prompt writing.

We think turning existing human video tutorials into machine-executable skills is a brilliant shift in methodology. That said, setting up headless software environments like Blender and local audio synths is still messy, and the extraction quality heavily depends on model reasoning. Still, for developers aiming to build autonomous agents that actually produce finished work in desktop apps, this is one of the most practical toolkits released recently.