Knowledge Skill

LLMs are amazing at summarizing content. Agents are amazing at doing workflows based on a simple input. I have found a huge amount of value by using OpenClaw as a "second brain."

I started by creating a knowledge/ directory and had my agent add artifacts into fixed directories (eg: "summarize this URL into the knowledge/technical directory.)

Over time I found I was endlessly adding more and more artifacts — a huge one has been summarizing YouTube videos into text, via youtube-dl.

Similarly the directory structure felt overly rigid — but elegant, so I was endlessly asking the agent to move files around.

Finally, I found that having all this context in the MEMORY.MD was a little cumbersome, so I made it into a skill so I could update and share it.

You end up with a skill that creates a directory like this:

knowledge/
├── CHANGELOG.md
├── README.md
├── books-to-read.md
├── fitness/
│   ├── _category.md
│   ├── raw/
│   └── summary/
├── parenting/
│   ├── _category.md
│   ├── raw/
│   └── summary/
├── reading-list/
│   ├── _category.md
│   └── books.md
├── tech/
│   ├── _category.md
│   ├── raw/  
│   └── summary/ 
├── unsorted/
│   ├── raw/
│   └── summary/
└── youtube/
    ├── _category.md
    ├── raw/ 
    └── summary/

Now that actual directories, such as fitness/ are fluid — the skill will move stuff around into logical directories as your knowledge grows. I asked Claude Code to outline the skill in a 3-5 bullet points:

  • Capture — Extract content from URLs, YouTube videos, files, or pasted text into raw files, auto-generate summaries, and pick up book references along the way
  • Organize — Sort entries into fluid categories, split/merge/rename categories as the knowledge base evolves, all confirmed with the user before execution
  • Recall — Proactively search the knowledge base when the user asks about any topic that might have been captured, using a lightweight memory index for ambient category awareness
  • Import — Onboard existing file collections into the knowledge base format with category mapping and dedup detection
  • Convention over configuration — Minimal frontmatter (title + source), filenames carry identity and date, sibling raw/+summary/ directories encode relationships, agent discovers cross-references at recall time
Visit project →