Connection Review

connection-review

Constraints

• NEVER scan whole vault; scope = 5-15 primary notes + small supporting set
• NEVER treat keyword overlap alone as evidence of connection
• NEVER synthesize week-level meaning (that is weekly-synthesis)
• NEVER write links without explicit user confirmation
• NEVER turn review into whole-vault graphing

Trigger Boundary

Use when: find note connections, suggest wikilinks, identify repeated patterns, surface contradictions, review relationships in a bounded set.

Do NOT use for:

• Raw inbox sorting → inbox-triage
• Week-level thesis and one action → weekly-synthesis
• Vault system health diagnosis → vault-health-feedback
• Upgrading material into stable notes → note-promotion

Boundaries

• Owns: link discovery, pattern detection, contradiction surfacing, follow-up suggestions
• Delegates to: note-promotion (follow-up accepted), weekly-synthesis (patterns span week)
• Never absorbs: inbox routing, synthesis, health diagnosis, context maintenance

Continuations

Companion Skill

All vault operations follow obsidian-mcp. Forbidden: obsidian_patch_note, obsidian_append_to_note. Verify every write through readback.

Procedure

Phase 1: Scope

1. obsidian_list_notes or obsidian_search_notes → identify recent/specified set
2. Bound to 5-15 primary notes; reject unbounded requests (ask user to narrow)

Phase 2: Read

1. For each: obsidian_get_note format:"document-map" → structure overview
2. format:"content" only when evidence is needed for a specific connection
3. Read older notes only to verify a suspected relationship

Phase 3: Analyze

When analyzing, read references/examples.md for connection quality calibration.

Four layers:

1. Link candidates: explicit note-to-note or note-to-project relationships
2. Pattern signals: repeated themes, problems, or methods across 3+ notes
3. Contradiction signals: new vs old claims, goal vs method, assumption vs evidence
4. Follow-up note suggestions: synthesis, topic, contradiction, or question notes

• Repeated method across 3+ notes = method pattern
• Same problem framed differently in 2+ notes = contradiction candidate
• Same entity referenced but never linked = link candidate
• Shared tag + temporal proximity alone ≠ connection (keyword trap)
• Contradiction requires real tension, not mere difference of topic
• A "pattern" with only 2 data points is a hypothesis, not a finding

Phase 4: Report

When formatting output, read references/templates.md for the exact report structure.

Default mode: analysis-only. Write links only after confirmation using obsidian_replace_in_note + readback.

Gotchas

Gotcha 1: Keyword matching as connection

What happens: Agent suggests links because two notes share the same tag or word Why it's wrong: Keyword overlap without conceptual relationship creates noise in the graph Correct approach: Require conceptual, project, or evidence-based relationship; cite specific content from both notes

Gotcha 2: Scope creep into whole-vault graphing

What happens: Agent starts reading notes outside the specified set Why it's wrong: Violates bounded scope constraint; wastes token budget exponentially Correct approach: Stay within 5-15 primary notes; read outside only to verify one specific suspected link

Gotcha 3: Review becomes synthesis

What happens: Agent extracts a thesis or "what it all means" from the connections Why it's wrong: That is weekly-synthesis territory; connection-review stops at relationships Correct approach: Report links, patterns, contradictions, follow-ups — do not interpret overall meaning

Validators (CI only, not runtime)

• validators/pre-check.sh: Confirms note set is bounded (≤15 primary notes)
• validators/post-check.sh: Verifies output contains no synthesis markers (thesis/emerging meaning)

Exit Criteria

• Every connection suggestion names both endpoints and cites evidence
• Patterns and contradictions are not based on keyword overlap alone
• Whole vault was not scanned
• Any written link has been read back and verified