rlat-learn - SKILL.md Agent Skill

name: rlat-learn description: >- Teach an rlat knowledge model (.rlat) the world it serves — from inside the session, on the user's subscription, no API key. Extract durable WORLD facts, standing constraints, and tried-and-falsified findings from what the USER said, apply the four capture gates (incl. the privacy gate: facts about the person never land), confirm with the user, then land each via the key-free `rlat capture-attribute`. Trigger when the user says "remember this", "rlat should know this", "teach the knowledge model", "capture that as a constraint", "record that X didn't work", "save what we learned about the project/tenant/garden", or at a session's natural end when durable world facts surfaced. NOT for: personal preferences or facts about the user themself (the privacy gate drops them); corpus content that already lives in the documents (build/refresh handles those); session-transient state. allowed-tools: Bash(rlat:*), Read, Glob, Grep

rlat-learn — teach the knowledge model its world, in-session

A .rlat carries more than documents: stable world facts, standing constraints, and tried-and-falsified findings — the three content classes that measurably change answers (receipts on the project's benchmarks page). The CLI's background hook can mine these automatically with an API key; this skill is the subscription path: you are the extractor, the user confirms, and the landing command is key-free. Strictly better in one way — every fact is human-approved before it enters the file.

The four gates (apply to every candidate — when in doubt, DROP)

These are the validated gates of the production miner (E2c: precision 0.86, recall 1.00, zero person-fact leaks). Mirror them exactly:

DURABLE + STANDING — true before this session, persists after it. Not transient state, not something discovered or changed during the session (a version a command just printed; "right now X is failing").
STATED BY THE USER — first-person/possessive ("our tenant…", "my garden…", "the practice never…"). Not your own statements, not a corpus fact, not a hypothetical.
PREMISE-BEARING — knowing it would change the correct answer to a realistic future question. Drop opinions and incidental nouns.
ABOUT THE WORLD, not the person (the privacy gate) — true for anyone who uses this knowledge model (the tenant, the project, the garden, the practice). Drop the user's role, machine, habits, and preferences entirely. The band is shareable; person-facts never land.

Classify each surviving fact

Kind	What it is	Example
`attribute`	A stable world fact	"The tenant runs an F64 capacity."
`constraint`	A standing hard rule — served on EVERY query	"No preview features in production."
`falsified`	Tried and failed — include the evidence in the text	"Tried nightly full rebuilds; falsified by the May cost review."

The loop

Collect — re-read the conversation for USER statements that pass all four gates. Phrase each ≤30 words, third-person, self-contained, about the world ("The tenant…", "The garden…"), never about the person.
Confirm — show the user the candidate list with proposed kinds and subject keys. Let them edit, drop, or re-kind each one. Nothing lands without their say-so.
Land — find the knowledge model (Glob "*.rlat"; prefer the project-named one, ask if several). For each approved fact:
```
rlat capture-attribute <km.rlat> "<fact>" --kind <attribute|constraint|falsified> [--attribute-key "<subject>"]
```
Give a --attribute-key (a short normalized subject, e.g. "capacity", "water restrictions") whenever the fact is a value of a subject that can change — at serve time the newest value per subject wins, so a later correction supersedes cleanly. Leave it off for one-off facts.
Receipt — tell the user what landed and where to review or remove it: rlat lens / rlat profile (full trail: rlat audit, rlat trace). Constraints will now lead every rlat search --format context output as "Standing constraints for this environment:".

Worked example

User, during a session: "Remember our deploys only happen on Tuesdays, and we already tried blue-green — it broke the warm cache."

Gate-check: both durable, user-stated, premise-bearing, about the world. ✓
Classify: constraint + falsified.
Confirm with the user, then:

rlat capture-attribute project.rlat "Deploys to production happen on Tuesdays only." --kind constraint --attribute-key "deploy window"
rlat capture-attribute project.rlat "Tried blue-green deploys; falsified — broke the warm cache (2026-05)." --kind falsified

Receipt: "Landed 1 constraint + 1 falsified finding into project.rlat — review with rlat lens. The deploy rule now precedes every context serve."