┌─ planner (opus) ─┐
   │ ui-agent (sonnet)│      QUERY / LINT
   │ builder  (sonnet)│ ─────▶ archivist (sonnet · read-only)
   │ reviewer (sonnet)│             │
   └──────────────────┘             ▼
            │                     VAULT
            └────── INGEST ──▶ scribe (haiku · writes)
                                     ▲
                                     │
                               Main Claude
                               (orchestrator)

What this is

A drop-in kit that extends Claude Code with:

• Two agents:
  • archivist — vault reader + health guardian (QUERY, LINT)
  • scribe — dedicated vault writer (INGEST, runs on haiku for efficiency)
• One skill — llm-wiki — the INGEST · QUERY · LINT protocols
• One constitution — vault-template/CLAUDE.md — the schema every agent reads before touching the vault (🇹🇷 Türkçe)
• One vault template — pre-structured Obsidian vault with conventions
• One linter — vault-lint.py, the LINT protocol as a runnable command (no dependencies)
• One retrofit guide — RETROFIT.md, for when you already have months of material
• Obsidian config — graph colours by layer, and the [!conflict] callout the skill needs

After install, your agents read from and write to a shared markdown knowledge graph that survives sessions. Yesterday's decisions are tomorrow's context.

The problem this solves

Default Claude Code agents are stateless:

Monday:  agents plan a feature, build it → close session
Tuesday: new session — "why did we use Stripe webhooks?"
Claude:  "I don't know. Let me re-analyze from scratch..."

Every session starts from zero. Context dies with the conversation. The project's reasoning evaporates even when the code persists.

With this kit:

Monday:  agents plan → scribe files to vault/decisions/
Tuesday: "why Stripe webhooks?"
archivist → reads vault/decisions/2026-04-21-stripe-webhooks.md → answers in 1 second with full rationale

Install

Quick

git clone https://github.com/selmakcby/claude-obsidian-kit ~/claude-obsidian-kit

cd ~/claude-obsidian-kit
./scripts/install.sh ~/my-project

Installs:

• ~/my-project/.claude/agents/archivist.md
• ~/my-project/.claude/agents/scribe.md
• ~/my-project/.claude/skills/llm-wiki/
• ~/my-project/vault/ (if doesn't exist)

Final step

Update your project's CLAUDE.md with vault permissions:

{
  "permissions": {
    "allow": [
      "Write(vault/**)",
      "Edit(vault/**)",
      "Read(vault/**)"
    ]
  }
}

How it works

The 6-agent team

After this kit, your full team:

The scribe / archivist split is token-efficient by design — see EFFICIENCY.md.

Three operations

🟢 INGEST (scribe)

Other agents produce findings. Main Claude passes findings to scribe. Scribe formats into proper note + writes to correct folder + updates index.

planner: "We should use monthly/annual pricing toggle"
  ↓
main Claude: Task(scribe, finding="...", type=decision)
  ↓
scribe: writes vault/decisions/2026-04-23-pricing-toggle.md

🔵 QUERY (archivist)

you: "Why did we choose monthly/annual toggle?"
  ↓
archivist: reads index.md → greps "toggle" → reads 2-3 notes → synthesizes answer

🟡 LINT (archivist)

you: "Check vault health"
  ↓
archivist: scans for orphans, broken links, stale notes, contradictions
archivist: returns severity-sorted report with fix suggestions

Retrofit — you already have a folder

vault-template/ assumes you are starting clean. Almost nobody is. The realistic situation is months of accumulated material — transcripts, drafts, half-finished projects, and often an older vault somewhere inside that already works.

RETROFIT.md is the guide for that case. It covers the two decisions you make before writing anything (the vault root is the folder you already have; an existing vault inside it is canonical and read-only), and the parallel-ingest design that keeps agents from writing the same entity page twice:

Phase 1  domain agents ──▶ project + decision pages only
                           (reference [[entities]], write none)
              │
              ▼  union + dedupe + deterministic split, in code
Phase 2  node agents  ──▶ disjoint slices of the entity/concept list
              │
              ▼
Phase 3  index + lint ──▶ one owner for index.md, one audit

Splitting by page type rather than only by domain makes collisions impossible by construction rather than by instruction. Reference run: 1.9 GB, ~1,000 text files, 226 pages, 3,484 links, zero broken links, zero fabricated source citations.

Lint, as a command

skills/llm-wiki/lint.md specifies the checks and a health-score formula. vault-lint.py implements that spec so the result is measured rather than asserted. Single file, standard library only, read-only:

python3 scripts/vault-lint.py path/to/vault

# retrofit case: audit the wiki layer, resolve links against the whole folder
python3 scripts/vault-lint.py . --notes 'wiki/*' --notes 'index.md'

# non-English frontmatter
python3 scripts/vault-lint.py . --fields tur,durum,guncelleme --date-field guncelleme

# CI
python3 scripts/vault-lint.py . --fail-under 85

Flagged > [!conflict] callouts are reported but not scored. A marked contradiction is the protocol working; the dangerous one is unmarked, and no linter can see that. Pass --score-conflicts for the strict formula.

Obsidian config

The skill tells the agent to flag contradictions with > [!conflict] — and Obsidian has no such callout type, so they render as anonymous grey notes. obsidian/ fixes that and colours the graph by layer. Copy it into your vault's .obsidian/.

Efficiency

This kit is intentionally designed for low token cost. See EFFICIENCY.md for all 10 patterns.

Highlights:

• Separate thinking from writing — opus plans, haiku files (80% cost reduction on writes)
• Index-first reads — never grep the whole vault
• Model tiering — haiku for scribe, sonnet for reads, opus for architecture
• Least-privilege tools — smaller system prompts, fewer wasted calls
• Parallel when independent — wall-clock savings

Rough cost per full feature (plan + design + build + review + 2 ingests): ~$0.18 Naive "opus everywhere" alternative: ~$0.90 → 5× cheaper with proper model routing.

FAQ

Common questions about INGEST, QUERY, LINT, PII, versioning, RAG comparison, and more — see FAQ.md.

Vault layout

vault/
├── index.md              ← entry point, read first
├── _schema/
│   ├── template.md       ← required frontmatter for all notes
│   └── conventions.md    ← naming, linking, tagging rules
│
├── architecture/         ← how the system is built
├── decisions/            ← why we chose X over Y (dated)
├── design/               ← UI/UX decisions
├── entities/             ← tools, services (stripe, supabase)
├── concepts/             ← cross-cutting ideas
└── sessions/             ← dated session summaries

Every note has frontmatter. Every note links to others. Orphans are tech debt.

File structure

claude-obsidian-kit/
├── README.md                 ← you are here
├── RETROFIT.md               ← pointing the kit at a folder you already have
├── EFFICIENCY.md             ← token patterns (10 of them)
├── FAQ.md                    ← common questions
├── LICENSE                   ← MIT
├── agents/
│   ├── archivist.md          ← reader/guardian (QUERY + LINT)
│   └── scribe.md             ← dedicated writer (INGEST)
├── skills/
│   └── llm-wiki/
│       ├── SKILL.md          ← main skill manifest
│       ├── ingest.md         ← INGEST protocol detail
│       ├── query.md          ← QUERY protocol detail
│       └── lint.md           ← LINT protocol detail
├── vault-template/
│   ├── index.md              ← vault entry point
│   ├── _schema/
│   │   ├── template.md       ← note template
│   │   └── conventions.md    ← full rulebook
│   ├── architecture/
│   ├── decisions/
│   ├── design/
│   ├── entities/example-service.md
│   ├── concepts/example-concept.md
│   └── sessions/
├── obsidian/
│   ├── graph.json            ← graph coloured by vault layer
│   ├── snippets/
│   │   └── vault-layers.css  ← [!conflict] callout + folder colours
│   └── README.md
└── scripts/
    ├── install.sh            ← one-command setup
    └── vault-lint.py         ← runs the LINT protocol for real (no deps)

Works well with

• Claude Code — the primary target
• Obsidian — the vault viewer/editor (optional — vault is plain markdown)
• VS Code, Typora, Bear — any markdown editor works
• Previous kit: knowledge-pipeline — this extends it with the multi-agent efficiency layer

License

MIT

Credits

Built on Claude Code's agent + skill system.

Inspired by:

• obra/superpowers — skills as methodology
• anthropics/skills — skill file format
• wshobson/agents — multi-agent orchestration patterns
• Zettelkasten — atomic notes + aggressive linking