narai-learnlisted

# narai-learn — put someone's habits into words they can hand to an agent narai records the moments its user corrected an agent. This skill reads that pile and writes the rules behind it. There is no API key in this. You are the model. `narai` supplies the corrections and throws out the rules whose evidence does not hold up. ## The loop ```bash narai score # how the last set of rules actually did narai corpus # what has been recorded # ...write rules.json... narai validate rules.json # exits 1 if a rule cannot be backed up narai validate rules.json --save # only once it exits 0 ``` Run `narai score` first and **open your report with how the previous set did** — which rules have had a correction of their kind arrive since, and which have not. If nothing has ever been proposed it says so in one line; move on. A distillation that never looks at the last one is just a fresh guess every time. Then run `narai corpus`, every time. Do not read the correction files directly — the corpus lays them out in a fixed shape, and the ids it prints are the ids `validate` will check against. It also lists **calls the user blocked**; those ids are citable like any other, but read them narrowly, since a block can be about one path or one moment rather than a standing rule. If it prints `nothing recorded yet`, stop and say so. Do not invent a starting set. ## What a rule has to be **Backed by at least two corrections, cited by id.** One