# LLM knowledge base skills
Agent skills that turn a directory of raw markdown notes into a self-updating knowledge base: tagged and backlinked notes, plus persistent wikis following [Andrej Karpathy's llm-wiki pattern](https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f).
You write things down; agents on a schedule handle the organizing.
## Install
```sh
npx skills add bholmesdev/llm-knowledge-base-skills
```
## The skills
| Skill | What it does |
| --- | --- |
| [enrich-note](skills/enrich-note/SKILL.md) | Enrich one note with topic tags (from a shared [tag registry](skills/enrich-note/references/tags.md)), source attribution, and links to related notes. |
| [enrich-notes-loop](skills/enrich-notes-loop/SKILL.md) | Run enrich-note across every un-enriched note in the vault, unattended. Uses an `enrichedAt` frontmatter stamp to skip finished notes. |
| [refresh-wiki](skills/refresh-wiki/SKILL.md) | Maintain every llm-wiki under `wikis/`: ingest new source notes, update entity/concept pages, and lint for stale claims and orphan pages. |
| [enrich-notes-loop-cloud](skills/enrich-notes-loop-cloud/SKILL.md) | enrich-notes-loop wrapped for scheduled cloud runs: sync an Obsidian vault down, enrich, sync back up. |
| [refresh-wiki-cloud](skills/refresh-wiki-cloud/SKILL.md) | refresh-wiki wrapped the same way for scheduled cloud runs. |
| [setup-oz-automations](skills/setup-oz-automations/SKILL.md) | Walks your agent through creating the [Oz](https://oz.dev) environment and schedules that run the cloud skills. |
## Expected vault layout
```
your-vault/
├── raw/ # raw, unfiltered notes — agents read these, never rewrite them
├── wikis/ # generated wikis — agents own this layer
└── tags.md # the tag registry — seeded by enrich-note on first run
```
Tweak the skills to taste if your layout differs.
## Visualizations
Two self-contained HTML apps that build views of your vault in real time. Drop them in the same directory as your notes and open the vault in [Hubble](https://hubble.md), which injects the file-reading runtime they use:
| File | What it shows |
| --- | --- |
| [notes-burndown.html](visualizations/notes-burndown.html) | A GitHub-style activity graph of your notes in `raw/`, with per-day tooltips. |
| [thought-constellation.html](visualizations/thought-constellation.html) | Your notes as a night sky: stars clustered by frontmatter tag (a payoff of enrich-note), related notes linked, click a star to read the note. |
## Running in the cloud
The `-cloud` variants are built for scheduled runners like [Oz](https://oz.dev). They assume the environment setup has synced your vault to `~/vault` with [Obsidian's headless CLI](https://obsidian.md/help/headless). Run the setup-oz-automations skill with your agent to create the environment and schedules — it pulls the skills straight from this repo, so there's nothing to push. Want to tweak the skills? Fork this repo and point the setup skill at your fork instead.
## Read the full walkthrough
These skills are [from my post on building a self-updating LLM knowledge base.](https://bholmes.dev/blog/llm-knowledge-bases/)