How I BuildThe Vault.
The Vault.
A workspace for
directing AI.
An Obsidian-based workspace where Claude and GPT do most of the work: persistent memory across sessions, specialized agents, quality gates, autonomous maintenance, and rules that get updated whenever something goes wrong. I prompt; the agents write code, notes, and content.
Core Methodology
Model Hierarchy
Opus architects and reviews. Sonnet implements everything. Haiku does research. Different cognitive loads, different costs. You don't use a $150/hr tool for $15/hr work.
Two-Pass Review
Sonnet implements fast. Opus reviews with fresh context — no sunk-cost bias. The implementing model defends its code. A fresh reviewer judges critically. Catches a lot of bugs before commit. Not perfect.
File Ownership
Parallelize by file, not feature. Each agent owns distinct files, so merge conflicts get rare. Multiple agents run in parallel; Opus reviews everything after.
The Vault
Everything runs through an Obsidian vault — a structured knowledge base that serves as the command center for all development. Agents read and write to it. Knowledge compounds across sessions instead of evaporating. The core principle: Markdown as the universal substrate. All state is plain text. Any tool can read it.
TestVault/
_collab/Agent protocols, knowledge, session state, heartbeat, inbox
protocol/ — RULES.md, AGENT-PROTOCOL.md, KNOWLEDGE-MAP.md
knowledge/ — Shared patterns, ADRs, swarm playbook, conventions
agents/ — 5 active role profiles (engineer, reviewer, vault-mgr, research, PA)
heartbeat/ 164 scheduled tasks, execution logs
loop-control/ — 6 session modes, active session tracking
Projects/10+ active projects, each with own CLAUDE.md
_scripts/692+ automation scripts (notify, search, index, heartbeat)
.claude/Skills, hooks, slash commands, agent definitions
Systems That Make It Work
Self-Improving Rules
Knowledge that gets smarter, not stale
14 critical rules promoted from real production incidents. Each rule carries evidence: date added, hit count, last verified, and source incident. Rules that go 90 days without being triggered auto-archive — the system fights knowledge rot by design.
The loop
Incident→CLAUDE.md→3+ occurrences→RULES.md→90-day decay
Example rule from RULES.md:
Auth on every mutation: getTenantId(ctx) as first lineadded: Jan 22hits: 12verified: Feb 16
Heartbeat System
164 autonomous tasks · Local GPU · Zero API cost
A systemd timer runs every 15 minutes, checking a schedule of autonomous tasks. All inference runs on a local GPU via Ollama with Vulkan backend (qwen3:8b and qwen2.5-coder:14b) — zero API cost. WARN/ERROR alerts push to Telegram immediately. Low-priority results batch into AM and PM digest messages.
02:00V3 Connect+Act (15-cycle burst — reads notes, implements improvements)
03:00Vault re-index (semantic search + knowledge graph)
05:00Vault backup + test runner
06:00Vault health check
07:00Morning briefing + daily context
08:30Fashion signals monitor
22:30Blog publisher
22:45Commit review
23:15Output quality review
Semantic Vault Search
15K+ notes · BM25 + embeddings
Custom-built vector search with chunk-level embeddings. Large files split by headings (markdown) or sliding windows (code) so no content is lost to truncation. Three search modes: semantic embeddings, BM25 full-text, and hybrid fusion via Reciprocal Rank Fusion. All local GPU, zero API cost.
Model
qwen3-embedding:0.6b
600M params · 1024 dims
Chunking
4,000+ chunks
Markdown by heading · Code by window
Search
Semantic + BM25 + Chunks
RRF fusion · NumPy vectorized
$ python3 vault-search.py "Convex auth middleware"
# Scores file + chunk embeddings, deduplicates by file, fuses with BM25
Multi-Agent Swarm Patterns
5 patterns · 10 rules from real failures
Five documented swarm patterns, each tested in production. 10 coordination rules extracted from real multi-agent failures — token waste, false positives, permission conflicts. Every failure became a rule. The system gets more efficient with each deployment.
ResearchImplementationQuality GateCounselDiagnostic
Feb 6: Audit swarm had 30% false positive rate → Rule 1: verification pass required
Feb 8: Diagnostic swarm burned 100K+ Opus tokens polling TaskList → Rule 7: leads yield, don't poll
Ongoing: Haiku + MCP permissions don't mix → Rule 10: use Sonnet for MCP-calling agents
Evolution
From first CLAUDE.md to autonomous vault maintenance in 7 weeks. Each phase built on the failures and learnings of the last.
1
Foundation
Jan 22 – Jan 26Jan 22Multi-tenant auth rules — first CLAUDE.md patterns written
Jan 24GiftTogether sprint — 9 tasks completed in 1 hour
Jan 25Two-pass workflow formalized (ADR 005). Vault deep clean.
Jan 26Parallel session system created — multiple terminals, one vault
2
Multi-Agent
Jan 27 – Feb 5Jan 27CRM features shipped. Swarm playbook started.
Jan 28Counsel review pattern — first documented multi-agent pattern
Jan 31Inventory Analytics Dashboard: 37 files, ~5k LOC, single session
Feb 5Buyer Order Form (12 files, 15 tests). Sentry integration.
3
Production Hardening
Feb 6 – Feb 10Feb 6Security hardening: 16 findings, 48 tests, email relay closed. Swarm playbook v1.
Feb 7-86-agent audit swarm benchmark. 10 swarm rules extracted from real failures.
Feb 10Production polish: 5 agents, 8 tasks, 3 deploys, 30+ items. Linesheet production-grade.
4
Vault Optimization
Feb 13 – Feb 18Feb 13Obsidian CLI integrated (11 operations). Progressive update protocol.
Feb 15Major cleanup: 4 dead systems archived, 5 agents retired, 6 profiles corrected.
Feb 17Semantic search built — 1,567 files indexed, 164ms queries, A/B tested.
Feb 18Local GPU sprint: ollama-rocm (15x speedup), all heartbeat tasks migrated. $0/day.
5
Shipping & Scaling
Feb 19 – Feb 23Feb 19Portfolio live at vachsark.com. 4 projects on custom subdomains.
Feb 20FreeFlow PDF: 6 free tools, full redesign. Linesheet demo-ready with seeded data.
Feb 22Linesheet buyer ordering: full order flow, share links, seller dashboard, email notifications.
Feb 23Vault search chunking (3,798 chunks). Local AI git hooks across 3 repos. Ordering docs.
6
Knowledge & Research
Feb 24 – Mar 13Feb 26Knowledge engine launched — automated multi-agent research pipeline seeding 700+ atomic notes
Mar 1Vault search v2: HyDE expansion, typed sub-queries, LLM re-ranking. Open-sourced.
Mar 5Research pipeline scored 0.5147 on the RACE benchmark in a one-off internal run. Real benchmark; unverified ranking against commercial systems.
Mar 9First fine-tuned local model deployed (vault-deepener-q8) — 10% to 100% format rate
Mar 1327 blog posts including Cross-Domain Isomorphisms series. Market Dashboard and Medicaid Analysis shipped.
Results
10
Production Apps
1
Developer
53+
Blog Posts
7 weeks
Foundation to production
Linesheet is in active use at my day job, managing 1,400+ inventory items. RepRewards has a handful of users logging workouts. FreeFlow PDF runs free tools client-side. These shipped from zero by prompting Claude — not by me writing code.
The thing that works for me is treating AI as infrastructure: persistent memory in the vault, quality gates that catch most (not all) bugs, specialized agents for different kinds of work, and rules that get updated every time something goes wrong. The vault gets a little smarter each session.
Want to talk about this approach?