Zettelkasten

Zettelkasten Skill

A personal knowledge base stored in SQLite. Atomic notes, linked together, searchable by keyword and tag.

All scripts live in workspace/skills/zettelkasten/scripts/ and use uv run. Database lives at workspace/skills/zettelkasten/db/zettel.sqlite.

First Run

Initialize the database (only needed once):

uv run workspace/skills/zettelkasten/scripts/init_db.py

Create a Note

Pipe JSON to stdin. Tags and links are optional.

echo '{"title": "Prompt injection has no solution", "body": "LLMs cannot distinguish operator instructions from instructions embedded in content.", "source": "Blog post", "tags": ["ai-safety", "llm"], "links": ["20260213170000"]}' | uv run workspace/skills/zettelkasten/scripts/create.py

Returns: {"id": "20260213180000", "title": "...", "tags": [...], "links": [...]}

Writing good Zettel notes

• Atomic: one idea per note
• Own words: restate the idea, don't just quote
• Context: explain why it matters
• Links: connect to existing notes where relevant
• Tags: use lowercase, hyphenated tags (e.g. ai-safety, event-sourcing)

Search Notes

uv run workspace/skills/zettelkasten/scripts/search.py "prompt injection" fts    # full-text search
uv run workspace/skills/zettelkasten/scripts/search.py "ai-safety" tag            # by tag
uv run workspace/skills/zettelkasten/scripts/search.py "20260213180000" id         # by ID (full detail + links)
uv run workspace/skills/zettelkasten/scripts/search.py "10" all                    # recent N notes

Update a Note

echo '{"id": "20260213180000", "title": "New title", "body": "Updated body", "tags": ["new-tag"]}' | uv run workspace/skills/zettelkasten/scripts/update.py

Delete a Note

uv run workspace/skills/zettelkasten/scripts/delete.py 20260213180000

Links

List links for a note

uv run workspace/skills/zettelkasten/scripts/link.py list 20260213180000

Add a link

echo '{"from": "20260213180000", "to": "20260213190000", "context": "Both discuss LLM safety"}' | uv run workspace/skills/zettelkasten/scripts/link.py add

Remove a link

echo '{"from": "20260213180000", "to": "20260213190000"}' | uv run workspace/skills/zettelkasten/scripts/link.py remove

Full graph

uv run workspace/skills/zettelkasten/scripts/link.py graph

Tags

uv run workspace/skills/zettelkasten/scripts/tags.py list                          # all tags with counts
uv run workspace/skills/zettelkasten/scripts/tags.py add 20260213180000 new-tag    # add tag
uv run workspace/skills/zettelkasten/scripts/tags.py remove 20260213180000 old-tag # remove tag

Embeddings

Generate embeddings for all notes (or re-embed a specific note). Uses OpenAI text-embedding-3-small (1536 dimensions). Requires OPENAI_API_KEY env var.

uv run workspace/skills/zettelkasten/scripts/embed.py                # all unembedded notes
uv run workspace/skills/zettelkasten/scripts/embed.py 20260213180000 # specific note

Run this after creating new notes to keep semantic search current.

Semantic Search

Search by meaning using sqlite-vec for native vector similarity (cosine distance). No numpy, no Python maths — all computed in C inside SQLite. Requires embeddings to exist.

uv run workspace/skills/zettelkasten/scripts/semantic_search.py "security risks of AI tool access"

Returns notes ranked by cosine similarity to the query.

Stats

uv run workspace/skills/zettelkasten/scripts/stats.py

Weeknotes Compilation

Compile notes from the past week, clustered by theme with connections:

uv run workspace/skills/zettelkasten/scripts/weekly.py        # last 7 days
uv run workspace/skills/zettelkasten/scripts/weekly.py 14     # last 14 days

Returns all notes, tag clusters (sorted by size, filtered for relevance), and connections between notes. Use this to draft weeknotes:

1. Run weekly.py to get the raw material
2. Identify 3-5 themes from the tag clusters and connections
3. For each theme, synthesise the linked notes into a narrative section
4. Include source attributions and links
5. Present as a draft with headings the user can edit

Trigger phrases: "compile my weeknotes", "weeknotes time", "what have I noted this week"

Workflow

When the user shares knowledge (from conversation, articles, videos, etc.):

1. Extract atomic ideas — one note per concept
2. Search existing notes for related ideas (semantic + keyword)
3. Create notes with appropriate tags
4. Link new notes to related existing notes
5. Embed new notes for semantic search
6. Report what was saved and how it connects

When the user asks "what do I know about X":

1. Search by keyword (FTS) and tag
2. Follow links from matching notes to find related knowledge
3. Synthesise an answer drawing from the connected notes
4. Cite note IDs so the user can drill deeper