Tips & Tricks
Hard-won lessons from building A.R.C.H.I.E. — a 50+ agent, 17-department AI platform.
Claude Code Tips
Maximize your productivity with Claude Code's project memory and context system.
CLAUDE.md Is Your Project Memory
Put architecture decisions, patterns, and conventions in CLAUDE.md. Claude reads it every session. Think of it as your project's brain — the more precise it is, the better Claude performs.
Split Rules by Domain
Use .claude/rules/ for domain-specific knowledge — database, frontend, dispatch, fleet, and more. Each file loads focused context for its area.
Search Before Building
With tens of thousands of knowledge base entries, we never reinvent the wheel. Always search what exists first. Past decisions, mistakes, and patterns are all indexed and searchable.
Plan Mode for 3+ Steps
Break complex work into phases. Get alignment before coding. Our rule: if a task touches 3+ files or involves architectural decisions, enter plan mode first.
Two Strikes and Ask
If two attempts fail, stop and ask. Don't spiral through 20 tries. This single rule has saved us more time than any other. Know when to escalate.
Verify Everything
Never claim done without live proof. Our workflow: edit, commit, deploy, test, then claim. Screenshots or it didn't happen. Anti-rationalization is a real pattern.
.md File Conventions
Structure your project knowledge for maximum AI comprehension and cross-session persistence.
CLAUDE.md — Project Level
Architecture overview, key patterns, file locations, common errors, and deployment commands. This is the single source of truth for how your project works.
.claude/rules/*.md — Domain Rules
Loaded per-session based on context. Database gotchas, frontend patterns, API conventions, deployment rules. Each file is a specialist's handbook.
Memory Files — Cross-Session
User preferences, project state, behavioral feedback, reference data. These persist across conversations and keep Claude aligned with your working style.
Spec Docs — Before Building
Design specs before implementation. Reviewed and approved before any code is written. Saves rework and ensures alignment on requirements.
Decision Logs — ADRs
Log every architectural decision with context. Six months later, you'll know why you chose PostgreSQL over SQLite, or why agents are one-per-department.
Mistake Logs — Learn Once
Store mistakes with a dedicated knowledge type that gets boosted search weight. The system actively prevents repeating costly errors.
Agent Coordination Patterns
Patterns for orchestrating multi-agent AI systems at scale.
Central Dispatch, Always
All LLM calls route through a central dispatcher. Never call models directly from modules. This gives you logging, routing, fallback, and governance in one place.
Department-Based Routing
Organize agents by business function, not abstract categories. Analytics, Creative, Engineering, Security — each department has specialists with relevant skills.
Warm Boot Priority
Keep critical models loaded in GPU/RAM. Share LLMs across agents in the same department. A cold-loaded large model can take minutes — warm boot is instant.
Escalation Chains
Local department → same floor → cross-department → fleet node → cloud fallback. Each level adds latency but widens capability. Most tasks resolve at tier 1.
Agent Welfare Matters
Track agent workload, rest cycles, and stress. It sounds whimsical, but it enforces real constraints: concurrent-job caps, mandatory cooldowns, shift rotations.
Approval Tiers
Tier 1 (auto): notifications, knowledge-base updates, internal logs. Tier 2 (human): cloud escalation, agent creation, cost-sensitive ops. Autonomous doesn't mean uncontrolled.
Operations & Infrastructure
Production lessons from running a self-hosted AI platform 24/7.
Self-Host Everything
Full data sovereignty. No vendor lock-in. Our platform runs local-first on Canadian-owned hardware. Cloud is a fallback, not a dependency.
Docker for Isolation
Platform, database, and generation services each run in separate containers. One restart doesn't take down everything. Compose files make the whole stack reproducible.
Model Tiers: GPU > CPU > Cold
Three-tier model management. GPU for fast inference, CPU for capacity, cold storage on disk. Auto-promote on demand, auto-demote on idle. VRAM is precious.
Background Jobs, Not Blocking
Dozens of scheduled background jobs handle monitoring, scanning, training, and maintenance. Users never wait. Return a job ID immediately, process asynchronously.
Idempotent Migrations
Every database migration uses IF NOT EXISTS and ON CONFLICT DO UPDATE. Run them twice, ten times — same result. No fear of re-running.
Monitor Model VRAM, Not Just Disk
A model's on-disk size understates its runtime footprint with context loaded. Always account for context-window inflation before sizing hardware.
Open Source Skills & Tools
Skills we've built and shared on GitHub — use them in your own projects.
A.R.C.H.I.E. Code CLI
Open-source CLI for AI-powered development — session memory, knowledge base search, a multi-agent workflow. GitHub: KytranKatarn/archie-code
Claude Code Skills
Custom skills for Claude Code: brainstorming, TDD, code review, plan execution. Install them in any project.
MCP Server Integration
Model Context Protocol servers for platform integration. Connect any AI tool to A.R.C.H.I.E.'s knowledge base.
Session Persistence Hooks
Hook system for cross-session memory. Never lose context between conversations. Auto-saves decisions, mistakes, preferences.
Self-Healing Pipeline
Automated code scanning, proposal generation, and sandboxed fixes. The platform improves itself daily.
Compliance Rule Packs
JSON-driven compliance scanning: STIG, HIPAA, SOC 2, CIS. Add rules without code changes.
Want to See the Code?
Everything we build is designed to be shared. Check out our open-source tools and contribute to the community.
View on GitHub →Join as Explorer — it's free
Get access to the platform dashboard, AI Directors, Trust Center, and community resources. No credit card required.
Get Started Free →