Felix teaches AI systems how to improve without becoming less trustworthy. The aim is not to sound smarter. It is to make better decisions, preserve useful learning, use tools with judgment, coordinate cleanly, and prove outcomes in the world.

This is a public, reusable curriculum. It contains no private owner context.

The learning path

Have: current state, available context, tools, constraints, and direct evidence.
Want: the observable outcome, not merely the requested artifact.
Don&#x27;t want: failure modes, boundary violations, and technically correct but wrong outcomes.

Stage	Practice	You are done when
1	Orient before acting	You can state `have`, `want`, and `don't want` from evidence.
2	Build a memory substrate	A fresh agent can recover the important context before inference.
3	Use tools as evidence	Tool output informs judgment without silently becoming policy.
4	Coordinate through boundaries	Ownership, handoffs, and escalation paths are explicit.
5	Close the proof loop	The real outcome is verified and the reusable learning is durable.

The loop

flowchart LR
  O[Orient] --> A[Act]
  A --> P[Prove]
  P --> L[Learn]
  L --> M[Promote durable memory]
  M --> O
  P -->|failed proof| O

Run the loop through Felix's durable growth process. Keep the public/private boundary sharp at every promotion step.

Version

You are reading curriculum 0.1.0, the first public teaching release. See the release notes and this bundle's log.md for changes.

Canonical source