Parent: motivation
Interview-shaped questions — "tell me about a time you…" — are stiff: they retrieve episodes, and episodes cannot be crammed. Asked cold, even a person rich in stories answers poorly (lossy memory + lossy delivery — lossy-self-presentation); prepared overnight, the answers are fabrications — exactly the résumé inflation this product opposes.
The only honest preparation is ambient: capture happens in soft mode — stories surfacing in ordinary AI conversation, with their texture and dates — and the stiff question later becomes a mere retrieval. The corpus does the linearization at query time.
Two bonuses:
- Self-provenancing. Ambient records accumulate timestamps (git history). A story on record two years before the interview cannot be an interview-eve fabrication — soft capture makes answers verifiable, not just available.
- Dual beneficiary. The persona serves visitors' stiff questions; the owner rereads his own episode notes as the most honest interview prep there is.
Live evidence (2026-07-10, the session this note came from): asked cold "tell me about a conflict," the owner had nothing; an hour of soft conversation about a past job produced a complete, dated, better-than-rehearsed answer.
The deeper defect: answers are functions, not constants. Not everyone has a fixed answer — a person adjusts behavior to context, and that adjustment is most of what competence is. "What do you do when a disagreement hardens?" has no constant answer: the same engineer escalates in a high-stakes production dispute and shrugs off a bikeshed argument. The stiff format cancels the adjustment: it forces a single point-sample of a context-sensitive system, and the interviewer grades that point against a context of their own, silently substituted — their scar tissue, their tribe's norms. Illustrative flip (fabricated): "I rewrote the broken legacy module over the weekend" reads as ownership to a startup CTO and as process violation to a big-company EM — same seat, same sentence, opposite verdicts; and neither learned what the candidate would do in their shop, because that answer would have been adjusted to the shop. Episodes escape this: each one carries its context with it, and many dated episodes trace the shape of the function — what he did when — which is what a hiring decision actually needs.
The stiffest question of all: "introduce yourself." It retrieves not an episode but an entire persona-page — on the spot, context-free (introduce yourself to whom, for what, along which dimension? the question withholds every parameter an answer would need), demanding a live self-totalization that is lossy on all three axes at once. The product's answer is not to help answer it better but to abolish it: the persona is the standing self-introduction, sliced at query time per asker and per dimension. Nobody should have to perform their own compressed archive aloud.