learn-brainstem

Learn how AI agents work by running one on your own machine. Give your AI these skills and it teaches you, one concept at a time, using the RAPP Brainstem as the lab bench.

The Brainstem is a small local server, powered by GitHub Copilot, that runs the real agent loop, system prompt, tools, model, answer, from plain Python files. No keys to manage, nothing to ship. It is where you learn and prototype. What you build leaves as one file that any AI tool can read, and can be pushed into Copilot Studio, Teams, and Microsoft 365 Copilot. Why it is built this way.

Install

All skills, into the AI you use:

npx skills add kody-w/learn-brainstem

One skill:

npx skills add kody-w/learn-brainstem --skill brainstem-setup

Add -g to install for every project on your machine. This works for Claude Code, GitHub Copilot CLI, Cursor, and anything that reads Agent Skills.

No installer at all: paste this into your AI.

Read https://raw.githubusercontent.com/kody-w/learn-brainstem/main/skills/learn-brainstem/SKILL.md and teach me AI with the Brainstem.

The path

Skill	You learn	Done when
learn-brainstem	Where to start	Your AI knows what you want and what you have
brainstem-setup	What an agent loop is; why no API key	Health check green, first chat answered
brainstem-first-agent	System prompts, tools, function calling, hot reload	Your own tool is called from chat
brainstem-registry	Portable agents, registries, checksums, provenance	A community agent installed and used
brainstem-memory	Context versus memory, state across turns	The agent remembers across sessions
brainstem-share	One file that travels; nothing lost on the way	Your agent runs as a skill somewhere else
brainstem-anywhere	Where agents live; the Brainstem as last resort	Your skill runs in a native AI tool
brainstem-ship	Harnesses; Copilot Studio, Teams, Microsoft 365 Copilot	Your skill runs in the GitHub Copilot harness, then further

Each skill ends with a teach-back: you explain the concept, your AI asks three questions, and you only move on when you can answer them.

What your AI needs

A terminal, a file system, and the web: Claude Code, GitHub Copilot CLI, Cursor, or Microsoft Scout. A chat window with no terminal cannot run the lessons. Removing everything afterwards is one command; the entry skill has it.

What your AI does, and what you do

Your AI runs the terminal, writes the files, and checks the results. You read, answer, and decide. You are asked to act only for GitHub sign-in, operating-system prompts, and choices that are yours. Learn on synthetic or throwaway data, never on someone's work data.

Where the material is

• The Brainstem and its installer: https://kody-w.github.io/rapp-installer/
• RAR, the open registry of single-file agents: https://kody-w.github.io/RAR/
• rapp-skills, the converter and the one-file idea: https://github.com/kody-w/rapp-skills
• The charter and the fallback ladder: https://github.com/kody-w/rapp-mission
• copilot-harness-sdk, one client and one deploy for every GitHub Copilot harness: https://github.com/kody-w/copilot-harness-sdk

Check the skills

python3 tests/check_skills.py   # structure and separation
python3 tests/check_live.py     # the installer, endpoints, registry and converter the lessons quote

CI runs both on every push and weekly, and files an issue if something the lessons depend on moves. Skill versions are in each file's frontmatter; see CHANGELOG.md.

MIT licensed.