Zettelkasten Linker

What this skill does

Takes a new note and a corpus of existing notes (Obsidian / Logseq / plain markdown) and proposes backlinks. Each suggestion includes the matched note, the relevant excerpt, and a similarity score.

Inputs

• new_note_path: the note being added.
• vault_dir: the root directory of existing notes.
• Optional top_k: number of suggestions (default 7).
• Optional min_similarity: cosine threshold (default 0.25).

Steps

1. Read the new note. Extract the title, tags (front-matter or #tag), and the body.
2. Tokenize the body, drop stopwords, lowercase. Compute a TF-IDF vector.
3. Walk vault_dir for *.md. For each existing note, compute a TF-IDF vector against the same vocabulary; cache vectors in <vault_dir>/.vectors.json for re-runs.
4. Compute cosine similarity between the new note and each existing note.
5. Take top top_k above min_similarity.
6. For each candidate, find the best-matching paragraph by re-running cosine over paragraph vectors.
7. Emit suggestions in a markdown block at the bottom of the new note (under ## Suggested Backlinks):
   • - [[<existing-note>]] — <one-sentence excerpt> (sim: 0.XX).
8. Detect orphan-prevention: if the new note has no existing backlinks AND fewer than 2 suggestions, recommend that the user manually link or place the note under a hub note.
9. Optionally inject [[wikilinks]] inline where the new note's prose textually matches an existing note's title.
10. Save the modified note (preserve original; write to <note>.linked.md unless --in-place).

Output

A modified note with a ## Suggested Backlinks section appended (or a sidecar file). Stdout prints the count of suggestions and a similarity histogram.

Verification

Open three suggestions and read the cited excerpt; the connection should be substantive (shared concept), not just shared stopwords. If the top suggestion's similarity is below 0.4 and the new note is short, the corpus may be sparse — recommend linking by tag rather than content. After accepting some links, re-run on the corpus and verify the new note now appears as a backlink-target for those existing notes' future revisions.

Edge cases

• Very short new note (under 80 words): cosine signal is weak; rely on tag overlap as a tie-breaker.
• Vault with thousands of notes: precompute vectors, then incrementally update only changed files (mtime check).
• Notes in mixed languages: vectorize per language, do not mix.
• Notes that are templates (e.g., daily journals): exclude from suggestions to avoid noise.