Course¶
A structured, self-paced course for building reliable AI agent sessions — from startup through work to shutdown. 8 modules, ~2.5 hours total. Each module builds on the previous one.
Foundations¶
-
Why unmanaged AI sessions waste tokens and produce unreliable results. Diagnose the specific pain points in your own project.
Time: 10 minutes
-
The 5-hook engine, manifest/sentinel state management, and 7 key design decisions. Build the mental model before writing code.
Time: 15 minutes
Hands-On¶
-
Build a working Level 1 system — SessionStart hook that runs infra checks and generates tier1 files from a YAML config.
Time: 20 minutes
-
Add structural enforcement — PreToolUse blocks tools until rules are loaded, UserPromptSubmit injects gate messages and catches infra FAILs.
Time: 20 minutes
-
On-demand tier2 loading with keyword triggers, drift detection with auto-heal, stop hook with retry loop, and session continuity.
Time: 20 minutes
Deep Dives¶
-
14 research-backed cognitive rules for faithful LLM outputs, organized into 5 phases to counter U-shaped attention bias.
Time: 15 minutes
-
Turn failures into structural enforcement: failure to learning to rule to audit check to hook. The system learns from its own mistakes.
Time: 15 minutes
-
Wire everything together for your project. Deploy the full system, run the smoke test, and verify end-to-end.
Time: 30 minutes
Prerequisites¶
- An AI coding agent installed (Claude Code, Cursor, Windsurf, Aider, or custom)
- A project you actively use an AI agent with (any language/framework)
- Basic familiarity with JSON and Python (reading, not writing)
Visual companion: Download the slide deck to follow along with diagrams.
Just want to deploy? Skip the course — run python3 setup.py and the
interactive wizard handles everything in 2 minutes.