r007b34r/opencode-obsidian-knowledge-workflow/connection-review
- community
Checking that the SKILL.md file is still in the repository.
Analyze relationships in a bounded set of Obsidian notes: candidate links, repeated patterns, contradictions, and follow-up opportunities. Load when user asks to find connections, suggest wikilinks, or surface contradictions. Do NOT use for inbox triage, weekly synthesis, or note promotion.
Install
npx vskill@latest install r007b34r/opencode-obsidian-knowledge-workflow/connection-reviewSKILL.md
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
| Condition | Next skill | |
|---|---|---|
| Follow-up note suggestion accepted by user | note-promotion | |
| Patterns span full week window, user requests synthesis | weekly-synthesis | |
| Notes too raw for relationship analysis | inbox-triage | |
| All connections weak/uncertain | Escalate to human |
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
obsidian_list_notesorobsidian_search_notes→ identify recent/specified set- Bound to 5-15 primary notes; reject unbounded requests (ask user to narrow)
Phase 2: Read
- For each:
obsidian_get_note format:"document-map"→ structure overview format:"content"only when evidence is needed for a specific connection- Read older notes only to verify a suspected relationship
Phase 3: Analyze
When analyzing, read references/examples.md for connection quality calibration.
Four layers:
- Link candidates: explicit note-to-note or note-to-project relationships
- Pattern signals: repeated themes, problems, or methods across 3+ notes
- Contradiction signals: new vs old claims, goal vs method, assumption vs evidence
- Follow-up note suggestions: synthesis, topic, contradiction, or question notes
<details> <summary>Pattern Detection Heuristics (expand when patterns are ambiguous)</summary>
- 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
</details>
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
Security
- Security scanPassed, 95 of 100Pattern checks on the skill's files.
- Model reviewPassed
Quality evaluation
- Task evaluationNot runNo evaluation yet. Run one in Skill Studio.
Source
- SKILL.md in the repositoryCheckingskills/connection-review/SKILL.md
Popularity
- Repository stars
- 12
- Repository forks
- 1
- 7-day trend score
- 13
Stars and forks are for the whole GitHub repository, not this skill alone. Not refreshed yet.
Works with
Any agent vskill installs to: 56 targets in all. These 39 have a page on this site.
- Claude Code
- Cursor
- GitHub Copilot
- Windsurf
- Codex
- Gemini CLI
- Cline
- Amp
- Roo Code
- Goose
- Continue
- Kilo Code
- Junie
- Trae
- OpenHands
- Qwen Code
- AI Agent
- Tabnine Chat
- Sourcegraph Cody
- Amazon Q
- Firebase Genkit
- Sweep
- Devin
- Phind
- Mentat
- Claude Desktop
- Aider
- GPT Pilot
- Plandex
- bolt.new
- v0
- Replit Agent
- Anse
- TypingMind
- BetterChatGPT
- LibreChat
- Warp AI
- Pieces
- Fig AI
Embed badge
[](https://verified-skill.com/skills/r007b34r/opencode-obsidian-knowledge-workflow/connection-review)