Field questions
FAQ.
Does it work with any assistant?
Yes — the method lives on the operator's side, not in a vendor's features: frame, signals, retake, tests, close. Interfaces differ in what they expose (visible reasoning helps), but the discipline transfers as is.
Do I need the protocol every time?
No. A request one answer can close is a consultation — the protocol would be over-engineering. The method pays where you'll return, dig, decide: long production, decisions, learning.
How fast does it pay?
From the first properly opened session: a frame changes the regime immediately. The deeper returns — signal-reading as reflex, clean closes, an externalized asset base — build over weeks of practice.
Does it prevent hallucinations?
No method does. It reduces their frequency and makes them catchable — sourced-or-marked claims, spot checks, the coverage test for long documents — but external verification stays your burden. Never zero.
What is the shift (la bascule)?
A rare, observable event in long well-steered sessions where production quality changes in kind, not degree. The book gives four types, conditions and dangers — stamped exploratory: repères, not results.
Why “session” rather than “context”?
Context names what the model is given; the session names what the two of you do across time. Context is one variable among many — the full articulation with neighboring terms is on the companion glossary at prompt-engineering.blog.
Is the book technical?
No prerequisites: no math, no API. It assumes only that you can open a chat window — and that you're willing to work: retake, dig, demand.
Where does the method come from?
From a dated 2026 field inquiry across some twenty architectures (matience.org), read through named instruments — Popper, Austin, Ginzburg — and formalized into this book. The contribution claimed is the articulation and the protocol, not the invention of the mechanisms.