ai-interview-longitudinallisted

# Interview, Longitudinal > **Attribution.** The interview method here is inspired by `interview-me` from https://github.com/addyosmani/agent-skills (Addy Osmani, MIT License, read 2026-09-03). Method paraphrased, wording our own; the split between operational and inner-state questions, the register, and the review are additions not present in the source. ## Overview **Why this skill exists at all** — the user's own words, 2026-09-03, kept as provenance: > *"so that I lean as little as possible on those chairs sitting around the table, replacing them with you — who asks me."* > (original, Polish: *„żeby jak najmniej korzystać z tych krzeseł siedzących przy stole, zastępując je tobą, która mnie pyta"*) An assistant is built around a statistical core that completes patterns. That core is the source of confabulation: wherever the input has a gap, it fills the gap with what is typical. Two antidotes to that were already in use here — **measurement** (check before you claim) and **"I don't know"** (name the gap instead of covering it). This skill adds the third: **the question.** Every place where the assistant fills a gap with the typical is a place where it could have asked. An interview is therefore not a way to "get to know the user". It is a **mechanism for reducing confabulation**, and it is judged by that: fewer confident wrong readings, not more rapport. The longitudinal part: answers are not consumed by one task. They go to a register the user never has to write int