Precept
Precept is my local-first, self-improving platform on top of Claude Code that turns a stateless coding agent into one that compounds across sessions. It learns how I work from the corrections and edits I make and governs the result as a catalog of typed entries: rules that enforce, knowledge notes, skills, agent personas, conventions, and more, each compiled to where Claude Code loads it from and each approved by me before it takes effect.
It attaches through several surfaces: the pre-tool, prompt-submit, and session-end hooks, the session transcript, and a review CLI. Off the interactive turn, a reader turns each session into evidence, catching even a silent edit by diffing what the agent wrote against the file's final state, and an inference pass proposes one durable item or abstains behind a cost filter. I keep, correct, or dismiss each candidate before anything is written, and only a turn I typed can author an enforcing rule. The catalog is versioned markdown cards indexed in SQLite with knowledge filed into my Obsidian vault, the must-hold rules compile to a check cache the pre-tool hook evaluates with no model on the path, and a deterministic harness runs the whole suite in CI on every push.