菜单
User-invoked: surface hidden concept connections via local embeddings (Ollama + nomic-embed-text). Flags related-but-unlinked (sentence, concept) pairs, then adds wikilinks where the connection holds. Invoke explicitly via /wiki-deep-link.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Answer questions from vault notes. Use when user asks 'how does X relate to Y', 'what do the papers say about Z', 'compare approaches to W', or discusses concept connections. Also fire immediately after wiki-ingest when user pivots from saving to asking how the new source relates to other vault topics. Also fire as a pre-ingest relevance check: before newly-shared content is dismissed as not vault-worthy, check whether it connects to existing vault notes. Also fire proactively as a conversation develops: when discussion drifts near vault topics, surface a genuinely non-obvious connection the vault holds — sparingly, only when it clears the bar. Do NOT answer from memory — read the actual notes.
ALWAYS use when the user wants a source saved to the vault — pastes an arXiv ID/URL, says 'add this to the vault', 'save this', 'remember this', 'I found this interesting', or describes a specific finding/technique/project to capture. Do NOT activate for general topic conversation — only when there's explicit save intent or a concrete source (URL, paper ID, pasted content).
Autonomous vault improvement loop (writes changes). Use when user says 'improve the vault', 'autoimprove', 'fix what you can', 'what's the vault score'. Do NOT use for read-only diagnostics — that's wiki-lint.
Fill concept stubs with citation-grounded definitions. Use when user says 'fill in the stubs', 'compile concepts', 'flesh out concepts', or as follow-up after wiki-lint reports stubs. Do NOT use for new sources not yet in vault — that's wiki-ingest.
Set up, list, rename, or merge research domains. Use when user says 'add a domain for X', 'set up a domain', 'what domains exist', 'change scope of X domain', 'merge X and Y domains'. Do NOT speculatively fire on ingest prompts — wiki-ingest hands off explicitly when no domain matches.
Deep paper breakdown (does not write vault notes). Use when user says 'analyze this paper', 'break down this paper', 'is this paper any good', or sends an arXiv ID/PDF with no save intent. Do NOT use when user wants the paper saved — that's wiki-ingest.
| name | wiki-deep-link |
| description | User-invoked: surface hidden concept connections via local embeddings (Ollama + nomic-embed-text). Flags related-but-unlinked (sentence, concept) pairs, then adds wikilinks where the connection holds. Invoke explicitly via /wiki-deep-link. |
| compatibility | Requires Ollama running locally with the nomic-embed-text model pulled. |
| disable-model-invocation | true |
Surface implicit concept connections that string-matching can't find. A paragraph discussing "iteratively adjusting parameters to minimize loss" should link to [[gradient descent]] even though the name never appears — this skill finds those connections.
Deterministic linking (commonplace link) handles exact name matches. But as the vault grows, agents traversing the knowledge graph hit blind spots where two things are clearly related but no wikilink connects them. This skill uses embedding similarity as a cheap pre-filter to surface candidates, then a Haiku agent reads the actual text and decides which connections to add.
This is not RAG. Nothing is retrieved at query time. This is a batch graph-improvement tool that strengthens the links future agents will traverse via wiki-query.
This skill requires Ollama running locally with the nomic-embed-text model. If not available, tell the user:
Deep linking requires Ollama with nomic-embed-text:
1. Install Ollama: https://ollama.com
2. Pull the model: ollama pull nomic-embed-text
3. Start Ollama: ollama serve
commonplace deep-link
Or scoped to a single note (useful after ingesting a new source):
commonplace deep-link --note "$VAULT_PATH/path/to/note.md"
Flags:
--threshold <float> — similarity cutoff (default 0.7, lower to 0.55 for more candidates at lower confidence)--top <int> — max candidates per note (default 10)The script embeds concept definitions and note sentences, compares them, filters out already-linked and name-mentioned pairs, and outputs candidates as JSON.
Present a summary to the user:
Deep link scan complete:
45 notes scanned, 82 concepts compared
2,400 text chunks embedded in 12s
47 candidate connections found across 18 notes
If zero candidates: "No hidden connections found — the vault's graph looks well-linked for the current concept definitions."
If candidates were found, dispatch the wiki-deep-linker agent with the candidates JSON and vault path inline. The agent reads the actual text of each candidate pair, applies its built-in linking rules, and adds wikilinks where warranted.
Agents have isolated context windows — they cannot see this conversation. Include the full candidates JSON and vault path in the agent prompt.
After the agent completes:
commonplace index --incremental
Report what was linked:
commonplace log --entry "## [$(date +%Y-%m-%d)] deep-link | Semantic linking pass\n- Scanned N notes, found N candidates, linked N connections\n"