Teaching it your preferences

An assistant out of the box has someone else's defaults. A chat product may remember some of what you tell it, but you cannot see what it kept, edit it, or know when it will apply. A box is one directory of your data, a card is one markdown file in it, and the agent is the coding agent that reads those cards before it acts, so what you teach it is a file you can read and edit.

What you do. Correct it when it gets something wrong. Answer the questions it asks. Say how you want to be spoken to: shorter, less cheerful, in Spanish. Say where a kind of thing belongs when it files something wrong. None of this is configuration you fill out in advance; it accumulates from use.

What the box does. A correction is written as prose at the place it applies, so the next item of that kind is handled the same way. A guide card holds the box's working theory of you for one kind of job: triage rules, actions, and reactions, each with a confidence level and a note of where it came from. A personality card holds voice and communication style, folded into what every agent reads before it acts; a briefing card holds a directory's situational context. Answering a question can write the fact into a guide, briefing, or personality card, so an answer becomes a standing preference. The box's own instructions file (CLAUDE.md, for the curious) and rules directory are the mechanism a coding agent already uses.

What it needs. Nothing beyond the box and some use over time. Making it yours, what it requires.

Where it is still rough. The rule sets do not rewrite themselves from a single low-confidence answer: turning a recurring pattern into a rule is a manual edit. The retrospective pass that mines recent chats for what you taught implicitly ships switched off. A box has one personality and one set of rules shared by everyone in it, so it cannot hold two people's preferences separately.

What makes it possible

Guides and personality (making it yours): a guide is a living document, a theory of you to be refined rather than a static config, which is what makes this teaching instead of configuring.
The questions loop (questions): every answer is precedent; the placement is the small part and the rule you just taught is the valuable part.
The history (durability and provenance): committing is what makes a change durable here, so what the box came to believe about you has a trail you can read.
Provenance (where a fact came from): your exact words are kept as a quote and a fact points at its origin, so what the agent inferred is always distinguishable from what you said.