Learn what is happening under the surface.
Each lesson adds one layer: the model, the agent loop, its reach, your safeguards, and the habits that make the work reliable.
- 01What an LLM is — and isn't
Tokens, context, training, inference, and why fluent answers can still be wrong.
- 02What makes an LLM an agent
The working loop: goal, context, action, tool result, revision, verification, and stop.
- 03Your first supervised task
Start small, set boundaries, watch the work, inspect the result, and keep control.
- 04Files, commands, tools & external actions
Classify what an agent can touch by mutation, scope, destination, privilege, and reversibility.
- 05Permissions, secrets & prompt injection
How sandboxes and approval modes work, what read access can expose, and when to stop.
- 06How to give a good task
Turn intent into a usable brief: goal, context, constraints, definition of done, and evidence.
- 07Review, test & verify
Separate a successful command from a correct result using diffs, tests, and acceptance criteria.
- 08Git, gently
Commits, branches, diffs, and recovery as a safety system for agent-assisted work.
- 09CLAUDE.md & AGENTS.md
Persistent instruction files, their scope and precedence, and why they guide rather than guarantee behavior.
- 10Skills, plugins & MCP
Reusable workflows and outside integrations—including the trust boundary each one adds.
- 11Cloud vs. local
Where models and tools run, what data moves, and the tradeoffs among privacy, cost, speed, and capability.
- 12Limits, model choice & evaluation
Choose by task and evidence, expect uncertainty, and build repeatable checks for quality.