By name: kg description: "Router for the Knowledge Graph + LLM Wiki system. Use when the user invokes /kg explicitly or asks 'which kg command should I use?'. Sub-commands — /kg-orient, /kg-update, /kg-query, /kg-lint, /kg-ingest (data ops) and /kg-reflect, /kg-challenge, /kg-connect, /kg-suggest (dialogue ops). Project-state auto-trigger (wiki/ or graphify-out/ presence) belongs to /kg-orient, not this router." trigger: /kg
kg — Knowledge Graph + LLM Wiki Router
A persistent, compounding knowledge base combining two layers:
- Graphify extracts structural relationships (god nodes, communities, surprising connections) from any corpus.
- LLM Wiki (Karpathy pattern) maintains a living Obsidian vault of entity pages, concept summaries, and cross-references.
The graph tells you what's connected. The wiki tells you what it means. Either layer works standalone — both together compound across sessions.
Activate When
- User invokes
/kg - User asks "어떤 kg 명령을 써야 해?" or how the Knowledge Graph + LLM Wiki system works
- User asks about the overall KG/Wiki architecture, ontology, or authority rules
Project-state auto-trigger (wiki/ or graphify-out/ presence at session start) routes to /kg-orient, not this router. See kg-orient/SKILL.md.
Do Not Activate When
This skill is a router, not an executor. Do not run an operation here — route to the corresponding sub-skill:
| User intent | Route to |
|---|---|
| 세션 시작 / 현재 상태 | /kg-orient |
| 지식 질의 | /kg-query |
| 새 문서 투입 | /kg-ingest |
| 그래프 갱신 | /kg-update |
| 위키 건강 점검 | /kg-lint |
| 패턴/긴장 발견 | /kg-reflect |
| 주장 반박 | /kg-challenge |
| 다음 자료 추천 | /kg-suggest |
| 암묵지 기록 | /kg-elicit |
| 시행착오 기록 | /kg-postmortem |
| 스키마 변경 | /kg-schema |
| 웹 조사 | /kg-autoresearch |
| Obsidian Canvas export | /kg-canvas |
| wiki 초기화 | /kg-init |
| 여러 프로젝트 통합 그래프 | /kg-merge |
| graphify MCP server 등록 | /kg-mcp |
Naming Convention
/kg-<verb>(hyphen) — canonical user-facing slash command. Always use this form./kg <verb>(space) — never used. The hyphenated form is ground truth in every sub-skill'strigger:field.
Common Rules (5)
- Raw sources are immutable. The LLM reads but never modifies them. They are the source of truth.
- Wiki and graph are derived artifacts. Both are regenerable views, never primary evidence.
- Verification chain: Graph → Wiki → Raw Source. Never cite graph/wiki as primary evidence for a decision.
- Human approval is required for: schema change, page reclassification, page deletion, source approval (autoresearch), heuristic promotion from Experience.
- Use hyphenated commands:
/kg-query, not/kg query.
References
For deeper detail, read the relevant reference file:
| Topic | File |
|---|---|
| System layout, source detection, page templates, technical notes, graphify v0.8.x CLI/MCP surface, freshness gate | references/architecture.md |
Fortran corpora (scientific/HPC code — libraries, solvers, legacy, earth-system incl. WRF) — .F/.f90 extraction, cpp step, node/edge model, USE/CALL graph workflow |
references/fortran.md |
| 7 classes, page_kind, instance_of, schema-as-product, proposal/receipt | references/ontology.md |
| Human vs LLM authority, hot.md priority, novelty test, usage rules | references/authority-matrix.md |
| BM25 search, hot/overview/_index hierarchy, archive policy | references/context-compression.md |
| Migration policy, supersession, convergence tracking, attractor test | references/schema-evolution.md |
| Codex ↔ kg bidirectional pipeline (review filing, domain context injection) | references/codex-integration.md |
Output Contract standard fields (Next command:, Confidence:, Caveats:, etc.) |
references/output-contract-standard.md |
Reference files are loaded only when needed (progressive disclosure). Sub-skills are self-contained and do not depend on this router for execution details.
Document Authority
- Sub-skills are the authoritative source for each operation's detailed flow.
- This router (
kg/SKILL.md) owns: routing table, common rules, references list. - Reference files own: ontology, authority rules, compression hierarchy, schema evolution, integration patterns.
- When sub-skill text diverges from a reference, the sub-skill wins for execution details; the reference wins for ontology/authority rules.