Mental Model A skill is a Markdown file that extends Claude’s behavior. Think of it as a persistent instruction set Claude loads on demand — not a script that runs automatically, but a detailed playbook Claude reads and follows when the right task comes up. How Claude finds and loads skills: At session start, Claude scans all registered skill directories and loads only each skill’s name and description (the frontmatter). Full content is not loaded yet.

Hardware TCO Sample Build: RTX 3090 24GB (used market, ~$700–800) Component Cost RTX 3090 (used) $750 Host server (Proxmox, used workstation) $400 32 GB RAM $80 1 TB NVMe $80 Power (600W system × 8h/day × $0.12/kWh) ~$21/month One-time hardware ~$1,310 Monthly power ~$21 Break-even calculation — you need to compare against API costs you’d otherwise pay. RTX 4060 Ti 16GB Build (~Budget) Component Cost RTX 4060 Ti 16GB $450 Mini PC / used workstation $200 16 GB RAM $40 512 GB NVMe $50 Power (250W × 8h × $0.

What Is the Shadow Stack? The “shadow stack” is a local inference layer that runs alongside your cloud API usage. Instead of every prompt hitting OpenAI or Anthropic, lightweight or private workloads run on GPUs you already own. You choose the right tier per task. Three deployment tiers: Cloud APIs — Claude, GPT-4o, Gemini. Highest quality, per-token cost, zero ops. Local inference — Llama 3, Mistral, Phi-3 on your hardware. Fixed cost after setup, full data sovereignty.