SKILL.md
Consilium-Principis — connect the server (onboarding)
Important, read first. This markdown skill is a funnel, not the product. The advisor logic — the fidelity contour (🔵 verbatim / 🟢 commentary / 🟡 extrapolation / 📐 calculation), fail-closed abstention, the quote gate, the council choreography, and every tool — lives in a local MCP server (scripts/mcp_server.py), not in this file. Installed on its own via npx skills add, this skill has no tools to call yet. Its whole job is to get you to the server.
Do not fabricate advisor answers or pretend the tools exist. If the MCP server is not connected, say so and walk the user through connecting it.
What to do
- Get the code. Clone or install the project:
``bash git clone https://github.com/ilyautov/consilium-principis ~/consilium-principis cd ~/consilium-principis python3 install.py # or run in place; Python 3.10+ ``
- Wire the MCP server into this agent. Generate a machine-correct config snippet:
``bash python3 scripts/board.py mcp-config --json # on Windows: py -3 ... ``
Then paste it into your host's MCP configuration. Host-by-host recipes (Claude Code, Claude Desktop, Cursor, Codex, Gemini CLI, Kimi CLI, and any universal MCP host) are in docs/CONNECT-HOSTS.md.
- Convene. Once the server is connected, say "where do I start". The server's
INSTRUCTIONS (delivered over the MCP initialize handshake) take over: it checks readiness (doctor) and seeds a starter board (seedcouncil: Marcus Aurelius and Epictetus, public-domain). If your host does not forward initialize instructions automatically, ask the agent to call the doctor and seedcouncil tools explicitly.
Why a separate MCP server
The moat is local execution with a bring-your-own private corpus: nothing runs on anyone else's infrastructure, and only public-domain content is ever packaged. A markdown-only skill cannot carry that guarantee — it can only point you at the server that enforces it.