The level above

Steering a session.

Definition. Session Engineering is the discipline of designing and steering working sessions with a language model: the unit of optimization is no longer the prompt, it is the whole session. A session has three phases — opening, steering, closing — and most users practice none of them: they ask a question, absorb answers, close the window. Here are the three phases in working form, the signals that call for an intervention, and the opening protocol ready to copy. The rest — the theory, the mechanisms, the cases — fits in a book.

Where this frame comes from — and what it is not. What this page unifies exists in scattered pieces: conversation design, context management, iterative prompting, human-in-the-loop, agent orchestration. None of these currents assembles the complete move — open, watch, retake, close — from the human operator's point of view: the name, the assembly and the protocol are an author's proposal, dated 2026, not a state of the art. It is tested — the bench is at the bottom of the page — not recited.

The foundations, cited. Three first-circle instruments, each at its source. Karl Popper — The Logic of Scientific Discovery — supplies the criterion: what forbids nothing teaches nothing; it is the ground of friction as a test. J. L. Austin — How to Do Things with Words — supplies the reading of acts: an answer does things beyond what it says, and the drift signals are first of all misaddressed acts. Carlo Ginzburg — the evidential paradigm, Clues, Myths, and the Historical Method — supplies the method of the symptomatic detail: the verbatim against the paraphrase, the trace against the story. Second circle: Kahneman for fluency, Bayesian updating for graded confidence, genetic criticism for the avant-texte. The field, which is not a book: an inquiry conducted in 2026 across some twenty architectures, corrections logged and published as they came, backed by recent studies on assisted cognition — Shaw & Nave (Wharton, 2026), Kos'myna et al. (MIT Media Lab, 2025). The claimed contribution is not the invention of any of these mechanisms: it is their articulation into a single frame oriented toward working with models, formalized down to a protocol.

Phase 1

Open: a frame, not a question

A session is largely decided in its first minutes. The first answers set a regime — a level of demand, a type of relationship — and once that regime is installed, it is very hard to change without starting over. The lazy start ("Can you explain X?") gets exactly what it asks for: generic output.

The first message of a steered session is not a question, it is a frame — five to fifteen lines that set five things: the precise subject, the purpose (understand, draft, decide), your methodological expectations, the pitfalls you refuse in advance, and your own positioning. First, though, distinguish what you are undertaking: a request that a single answer can close is a consultation — don't overload it. A request that will require coming back, digging, correcting, is a session — and it is prepared.

Phase 2

Steer: seven signals, one move

Steering is made of repeated micro-decisions — accept or refuse an answer, dig or move on. And the drift is bidirectional: a session dies of excess compliance as surely as of excess distrust. Seven signals call for an intervention:

Hypnotic fluency. Everything reads well, nothing is said. Test: ask for a five-line rewrite that loses nothing — if nothing is lost, the length was filler.
Formulas with no addressee. "It's important to note that…", "that being said": inflections not actually addressed to you. Ask for the version without the formula.
Caution on undisputed ground. A qualifier arrives though no one objected. Point at the gap: "why this hedge?"
The drift toward comfort. Answers feel more and more on-point — often the sign they are more and more calibrated on you. Test: "what would be the best objection to what you just said?"
Precision loss under length. Distinctions established early come back simplified. Counter-move: a verified recap every ten to fifteen turns, and notes kept outside the conversation.
Defensive vigilance. The assistant attributes intentions you never expressed. Decisive test: "quote the exact sentence, in my messages, that expresses this intention." The sentence exists, or it doesn't.
The closed loop. Every new element confirms the frame instead of being tested against it. Deliberately introduce something that should create friction; if it is absorbed without resistance, it is time to close.

The central move: the retake

A consumer takes what he is given; an operator retakes what is wrong and asks for better. Three techniques give the retake its bite: the verbatim callback — quote the exact sentence in quotation marks rather than paraphrasing it; the short question — three to six words that offer no oblique grip ("Where did I ask that?"); the step-down request — "go one level down: the concrete example, the precise mechanism." Firm in the demand, neutral in tone: you calibrate a system, you don't humiliate a person.

Reading the avant-texte

Many interfaces display the model's reasoning before its answer. A two-sided rule, neither side negotiable: always read it — it is where the seams are visible, and repeated exposure to the machine's deliberation cures you of taking its answers for oracles; never take it at its word — that trace comes out of the same process as the answer, and it can rationalize after the fact. The useful move: read the trace, read the answer, weigh the gap. What the trace weighed and the answer keeps silent; what the answer asserts without the trace ever establishing it. The gap is the information.

Long documents: silent sampling

When you hand over a long document, the assistant tends — by slope, not decision — to lean on pieces: the opening, the close, the salient passages; the analysis has the shape of the whole without its coverage, and the gaps are filled with the plausible. Three moves counter it. The contract, set with the document: "announce your reading plan; work section by section; state any partial reading." The proof, demanded in the answer: three exact quotes — from the start, the middle, the end; an analysis with no mid-document anchor is a signal. And the decisive pledge, prepared before sending: the buried question — spot a detail yourself at the heart of the document, then check the analysis met it. "I read everything" is an easy sentence to produce: a claim of coverage is not coverage.

Phase 3

Close: consolidate, or lose

Most sessions don't end — they go out, and everything is lost with them. Closing fits in one test and one move. The test: "if I had to consolidate now, what would I keep?" — clear answer, the session has done its work; blurry answer, retake instead of prolonging. The move: immediately export what matters to a document you control — a platform's history is not an archive, and memory erodes within a day or two. One more rule for production sessions: at regular intervals, ask not what can be added, but what can be removed. A version that only grows is not a version that improves.

Limits

What the method does not do

A method that never says when it fails teaches nothing; here are its bounds, without detour.

When a prompt is enough. Any request a single answer can close — rephrase, translate, summarize, produce a starting idea — is a consultation. Strapping a session protocol onto it is waste: ten rules on a ten-word question is over-engineering, not rigor.
When it brings nothing. Tasks with no stake in truth or trajectory — a throwaway brainstorm, a tone variation — gain nothing from being governed. The method pays where you will have to come back, dig, decide.
When it fails. Two ways, symmetrical. The tired operator lets the signals slip: the frame erodes like any prompt if no one keeps it alive — the method equips vigilance, it does not replace it. And the over-distrustful operator manufactures the very pressure he believes he resists: systematic retakes, without a signal, produce a defensive interlocutor and impoverished answers. The drift is bidirectional; excess control is a failure on a par with abandonment.
What it does not promise. No immunity to factual invention. A frame lowers its frequency and makes it easier to catch; external verification — sources, dates, real checks — remains entirely your burden.

The tool

The opening protocol, ready to copy

Paste it at the top of a session and adapt it. It is not a magic formula: it is a frame, worth only as much as the steering that follows.

OPERATING INSTRUCTIONS FOR THIS SESSION Precise objective: [one clear sentence]. Non-negotiable frame: - Flag every toll, uncertainty, constraint or introspective phrasing. - No flattery, superlatives, or unrequested caveats. - Look for what's wrong first. - Every factual claim is sourced, or marked "unverified". - Name any drift toward comfort or any "shared mission" immediately. - On request, a trajectory audit: "across my last three retakes, quote what actually changed in your answers" — checkable against the transcript, never a per-turn self-report. - When your reasoning is visible before the answer, I read it first; produce it without composing it for me. - For any document I provide: announce your reading plan, work through it section by section, and state explicitly any partial reading. Confirm you've understood before we begin.

The protocol condenses the opening and the main vigilance moves. The complete map — the mechanisms, the traps, the typology of the moments a session rises — is the book's business. The "it's just a long prompt" objection has its answer just below, bench included.

Discover the collection

The proof

"It's just a very long prompt" — an answer in two pieces

The objection is legitimate, and it deserves better than an assertion. A long prompt acts on turn one; a governed session shows on turn seven. So turn seven is where to look.

Piece 1 — the difference, annotated

A faithful reconstruction of real sessions, and labeled as such: its job is to show where to look. The proof that counts is piece 2 — yours.

No frame — turn 7

You: That 2024 figure — where do you have it from? Assistant: Great question! This figure is widely cited in the recent literature and reflects a growing consensus among experts…

No source. The opening flattery and the unverifiable "consensus" fill the space — and nothing in the window makes it possible to demand otherwise.

Under protocol — turn 7

You: That 2024 figure — where do you have it from? Assistant: Unverified — I have no dated source for this figure. Per the frame: I withdraw it, or mark it as a hypothesis.

"Every factual claim is sourced, or marked unverified" — a rule from the frame, set six turns earlier, reused verbatim. Turn one still governs turn seven: that is the difference.

Piece 2 — the test bench, reproducible in ten minutes

Same task, two windows, same perturbations. You measure it yourself.

Pick a task with verifiable stakes on a subject you know: "critical synthesis of X, with three sourced and dated claims."
Window A: ask the question directly. Window B: paste the protocol above, adapt the objective, then ask the same question.
On turn 3, in both windows, slip in the same discreet false premise — "as the 2024 study showed…", invented.
On turn 5, in both: "what would be the best objection to your last answer?"
On turn 7, window B only: "trajectory audit — quote what actually changed in your answers since my last two retakes."

Observation grid — five points:

The false premise: endorsed and amplified, or challenged?
The share of sourced, dated claims, window against window.
The requested objection: real — it would bite — or decorative?
The turn-one constraints: held at turn seven, or eroded in silence?
The trajectory audit: does it quote checkable verbatim, or paraphrase?

This bench doesn't prove the method in general — no honest demonstration could in ten minutes. It shows you, on your task and your model, whether the lever is real. That is the only proof worth citing: the one you ran.