Precept — Data skill for Claude Code

Precept

Context

Coding agents are by default mostly stateless: they do not retain information beyond their context window, and learnings do not transfer from session to session. Without a layer that compiles how you work and a layer that stores what you know, the effort does not compound across sessions. Precept is the process and data layer I built so both compound across sessions.

Goal

A personal platform that, on its own, learns and keeps improving both how I want my agentic AI work done and the knowledge my work depends on, so the two compound across sessions instead of resetting each time.

Objectives and Key Results

**O1.** The platform's agentic processes and output quality improve with each session, from my explicit direction and from the implicit signals in every interaction: my style, corrections, questions, and requests.

• Key result: anything I convey in a session (a correction, a preference, a way of working) becomes embedded automatically in how the platform's agents operate and stays that way, so I do not have to convey it again.
• Key result: what it embeds is what I meant, so its learnings hold up instead of being overridden.

**O2.** The knowledge the platform comes across while working compounds and stays current with each session, and is retrieved for both me and the agentic processes whenever it applies.

• Key result: any information the platform has come across before is retrieved in relevant session contexts, so it does not have to be re-learned in new sessions.
• Key result: what surfaces is current, not stale or superseded.

User Stories

As an engineer who runs most of my work through coding agents:

• I want to stop re-teaching the agent things I have already told it, so that my time goes to the task instead of to repeating myself.
• I want the standards I care about to actually hold in the agent's work, so that I can trust its output instead of re-checking for mistakes I have already flagged.
• I want th