/relink — wire the important stuff into the graph (don't just add a note)

🧒 When reporting to a non-technical operator: end with a child-simple "In plain words" recap in their language.

The principle (canon: concept-creation-rules.md §11): an important new idea that only exists as a note is an island, not part of the Second Brain. It must be integrated into the knowledge graph. Mirror of [[no-orphan-notes-rule]] (passive "≥1 inbound"); this is the active "maximize meaningful connectivity for important nodes" side.

Architecture law (from the DR): retrieval first → judgment second → writing last. The LLM never decides links by reading the whole vault; it gets a bounded, diverse candidate pool, then types relations and proposes edits. Less hallucination, explainable, repeatable.

What counts as "important" (don't run for every note)

A NEW: concept · mental model · framework · theory · term · project · philosophy · research cluster · personal principle · knowledge-organization system. Threshold = §1 of concept-creation-rules (≥3 recurrences, noun-entity, domain-bound). Routine notes → passive no-orphan check, not this.

⚠️ Token law — channels FIND, LLM JUDGES (never scan all 154k notes)

The operator's law ([[vault-data-architecture]]): cheap tools first; the LLM judges only the top-K.

---

Mode A — integrate ONE new node (default)

Input: the new note's path/name (or freshly pasted text → first save via obsidian-ingest, then relink).

1. Extract entities (cheap — read ONLY the new note)

List: canonical concepts · aliases/acronyms · key claims · related domains · possible parent/child concepts · candidate MOCs. (LLM + the note's own frontmatter; no vault scan.)

2. Gather candidates from 3 channels (NOT scan-all)

• Lexical — python "$IMPORTS_ROOT/namesearch/find_name.py" "<each named entity>" (exact/translit/typo, 0 tokens) + grep literal mentions + alias/unlinked-mention matches.
• Semantic — python "$IMPORTS_ROOT/brain_ask.py" "<node + its entities>" (e5+reranker; finds hidden 2nd/3rd-order links: causality, analogy, shared mechanism, opposing view). e5 is multilingual → RU/EN drift is covered.
• Hub-awareness — incoming_counts in $IMPORTS_ROOT/orphan-scan/reverse-index.json (basename→inbound count, free): mark which candidates are hubs/MOCs (high count) vs orphans (0). Used by the reverse-link + MOC policy below. (Full 1/2-hop adjacency graph = deferred future enhancement; semantic channel already covers conceptual neighbours.)
• Cross-check 06-Concepts/ + aliases + 09-Bridges/ so you don't propose a dupe (defer to concept-creation-rules §5).

3. Fuse + rank P0–P3

Fuse the channels (RRF-style — don't add raw scores from different rankers). Soft ranking heuristic: ≈ 0.35·semantic + 0.20·lexical + 0.15·entity_overlap + 0.20·hub/MOC_fit + 0.10·freshness (zero freshness if evergreen).

• P0 — missing link breaks navigation/understanding; exact alias/unlinked mention; parent concept; canonical MOC missing; evidence from 2+ channels.
• P1 — strong thematic / cross-domain bridge / method / example / contrast.
• P2 — useful "see also"; add only if it doesn't bloat the note.
• P3 — weak/speculative → DON'T write; log in the memo as backlog. No link without evidence. Low confidence → defer, don't write.

4. Type each accepted link (closed vocabulary)

defines · defined_by · extends · depends_on · contrasts · example_of · evidence_for · method_for · same_cluster · moc_member · see_also. Write it as a short gloss after the link: - [[Target]] — method_for: <one phrase why>.

5. Build the change-set (PREVIEW — don't write yet)

For each edit produce an internal record {path, target_type: heading|frontmatter, target: "Related", operation: append_once, text, dedupe_key: "source::target::relation"}.

• Forward links — into the new note's ## Related (always for P0/P1, sometimes P2).
• Explicit reverse links — SELECTIVE, not for symmetry's sake. A forward link already creates an automatic backlink, so add an explicit reverse inline link ONLY if: target is a concept hub/MOC · relation is asymmetric & operationally important (depends_on/defined_by/extends) · target loses context without it · source gives target a new example/contrast/method. Do NOT add a reverse link into a dense hub/glossary where the auto-backlink already suffices (avoids overlinking).
• MOC/hub wiring — this is the real "old→new" that matters: ensure the new node appears in the right MOC/hub (≤2 MOCs; new MOC only in --deep). Prefer MOC sections Core/Related/Methods/Debates/Examples over a dump.
• Missing notes — propose; create concepts yourself per concept-creation-rules §1 (no-ask).
• Caps: ≤5–7 new links per normal evergreen note per run (MOCs exempt), ≤3 per section.
• Show the change-set as a BEFORE→AFTER table on the operator's REAL notes ([[show-before-after]]). Wait for their OK.

6. Backup → apply (idempotent)

• python "$IMPORTS_ROOT/vault_backup.py" BEFORE any write ([[vault-backup-rule]]; runbook = skill obsidian-backup).
• Apply each change-set record with append_once semantics: before adding a link, check it isn't already in the section (dedupe_key) — re-running /relink must be a no-op on already-done links.
• Insert into the ## Related block; never mangle adjacent list lines (⚠️ dedup-skill grabli: never batch-Edit list deletions).
• Optionally stamp relink_last_run: <date> in the new note's frontmatter (feeds the monthly audit).

7. QA → reindex → report

• Preflight/QA (blocking): broken target, duplicate patch, missing evidence for a P0/P1, too many links in one section, MOC update without a relevance reason.
• python "$IMPORTS_ROOT/validate_links.py" → 0 broken.
• orphan re-check on touched notes; python "$IMPORTS_ROOT/brain_embed_update.py" → RAG sees the new edges ([[reindex-routine]]).
• Integration Memo to the operator: importance reason · main bridges (typed) · notes updated · links added · MOCs touched · concepts created · deferred/rejected candidates · QA counts. End with 🧒 recap.

---

Mode B — --deep (deep relink: whole vault → islands)

Monthly/on-command sweep over the ALREADY-COMPUTED orphan list (not a fresh full scan).

1. python "$IMPORTS_ROOT/orphan-scan/orphan_scan.py" → refresh orphans.csv / orphans-by-folder.csv ([[vault-orphan-baseline]]).
2. python "$IMPORTS_ROOT/orphan-scan/build_dashboard.py" → $OBSIDIAN_VAULT/_Dashboards/Vault-Orphans.html (the operator works by eye).
3. Also surface: orphan clusters · missing MOC membership · alias gaps · high-centrality notes (high incoming_counts) without a hub · stale MOCs. Run top island clusters through Mode A steps 2–7. Loop-until-dry (stop at intentional orphans: archives, raw originals).
4. Dupes/near-dupes found → hand to skill dedup (don't merge here).
5. New MOC may be proposed here for a stable 5+ note cluster with no hub. Report what got woven in + what was deliberately left.

Mode C — relink a whole THEME / topic cluster ("relink everything about X")

Between A and B: the operator names a TOPIC, not one note. Run Mode A steps 2–7 with the THEME's entities as the query (RAG + namesearch + grep on theme terms → judge top-K). Then, before proposing links:

• Map what already exists (concepts + MOCs + bridges in the domain) — don't assume it's empty; mature themes already have a rich cluster.
• ⚠️ Watch for FRAGMENTED HUBS — the #1 finding of a theme run. A long-lived theme often grew 2–3 rival hub/MOC notes that don't cross-link (e.g. a personal-POV concept + an import-corpus MOC + a thematic MOC). Pick ONE umbrella hub (usually the 90_MOCs/ one), wire the others up to it, and roster the sub-concepts under it. This consolidation is the highest-value edit — bigger than any single link.
• Don't re-link an already-dense sub-cluster (siblings already cross-linked) — only its membership in the umbrella + auto-backlinks. Avoids overlinking.
• The recurring signature to hunt FIRST: THREE LEGS of fragmentation that don't cross-link. Confirmed 5× (archaeology · health · PhD · dedup · father). The three legs of any personal-domain cluster:
  1. Import-corpus MOC — holds the operator's OWN raw data (TG export, email archive) e.g. _Health-MOC, _Perepiska-s-Ottsom-MOC
  2. Thematic 90_MOCs/ MOC — holds external/field knowledge, parent MOC-index e.g. MOC-Biohacking-Longevity, MOC-Family-Life
  3. Identity-layer _Self-Bible-MOC — holds the "this is who I am" view (cofounder of digital twin) e.g. PhD-as-scholar, family-as-influence
• A long-lived theme typically has the body of one leg but is invisible from the others. The highest-value Mode C move = wire all three together (+ make sure root MOC-index lists the import-corpus too, not just the thematic).
• Don't re-link an already-dense intra-cluster (overlinking guard).
• Proven runs:
  • 2026-06-14 (archaeology/alt-history): 3 fragmented hubs unified, research-swarm argument map woven into the concept layer — 7 notes / ~30 typed links / 0 new broken.
  • 2026-06-16 (health): personal _Health-MOC (830 notes) was orphaned from thematic MOC-Biohacking-Longevity + MOC-index — bridged both ways, 6 personal concepts rostered, 5 concepts up-linked — 8 notes / ~16 typed links / 0 new broken.
  • 2026-06-20 (PhD): minor — utility morning consolidation had already done ~90%; only 3 hub edits needed (MOC-index + cross-domain bridges to crypto/AI) — 3 notes / ~3 links / 0 broken. Lesson: always verify-existing first; PhD was already 90% done.
  • 2026-06-20 (a close family member): all three legs missing in the integration — created a person-card person-<name> + MOC-Family-Life (closed long-standing TODO in MOC-index) + identity link in _Self-Bible-MOC (father = "identity-brick of the twin"). 5 notes / ~12 typed links / 0 broken.

Safety (hard gates)

• Bidirectional = WRITING into old notes → backup before, preview before, never auto-write without the operator's OK.
• Never delete; never blanket-glob; never pattern-touch concept-*/person-* ([[operating-agreement]], [[vault-conventions]]).
• New concepts: create per concept-creation-rules (no-ask per §1); the LINK plan into old notes is shown for approval.

Rejected complications (don't re-pitch — see [[relink-mechanism]] / [[declined-decisions]])

The DR suggested, and we declined for AK-47: a Local REST API + MCP plugin (Edit targets sections fine), a new vector DB (have e5+reranker), spaCy 3-layer NER (LLM+namesearch suffice), a JSON-schema validation runtime, and a full link-adjacency graph engine (hub counts are enough for v1).

See also

• concept-creation-rules.md §1 (when) + §11 (integration) — the canon.
• skill ask (RAG engine), dedup (dupes), obsidian-ingest (first-time save of a raw dump), obsidian-backup (backup runbook).

<!--kit-footer-->

---

Like this skill? It is one of 100 in second-brain-starter-kit: the second brain we built for ourselves and run every day at Palo Alto AI Research Lab. Install the whole set with npx skills add tonydzi/second-brain-starter-kit. Everything is open source and free, so take what you need.

Flagships worth a look on their own: secondop-panel (a second opinion from a panel of external models), claude-memory-tidy (stop your agent's memory from rotting), telegram-mcp-kit (your own Telegram over MCP in about 15 minutes).

Author: Anton Dziatkovskii, Palo Alto AI Research Lab. Telegram @tonydzi - WhatsApp +1 341 222 9178 - X @Tony_Stef_

Engineers: want to test-drive this setup? Message me. I hand out free starter seeds to engineers who test and report back, and custom skill requests are welcome.