The terms, in their places

The articulation glossary.

The vocabulary of model interaction is abundant, and every term is excellent at its own level. This page doesn't pit them against each other — it articulates them. A fairness clause, first: the definitions below are working definitions, loyal to how each field uses its own name; if a practitioner of any of these domains finds one unfair, the correction is welcome — a map must not flatten the territories it borrows. What changes here is none of the terms: it is the unit of analysis.

prompt → answer.
context → answer.
session → trajectory → result.

Entries

Ten entries, four beats each

Session Engineering

What it names: The discipline of designing and steering working sessions with a language model — opening frame, drift signals, retakes, reading the avant-texte, consolidated close — with the human operator as the permanent control layer.
Its strength: This site's proposal (2026). The claim is not a new technique but a new unit of analysis.
Where it lives in the session: It is the session: everything below lives inside it.
Seen from the session: The shift itself: from prompt → answer and context → answer to session → trajectory → result. Every neighboring term keeps its excellence at its own level — and acquires an address.

Prompt engineering

What it names: Crafting the individual request: precision, purpose, examples, constraints, positioning.
Its strength: Single-shot quality; repeatable templates; getting the most out of one turn.
Where it lives in the session: The opening's raw material, and the phrasing of every turn.
Seen from the session: It becomes a local sub-discipline. Phrasing still matters at every point — but the point is not the unit, and a perfect prompt cannot hold its own constraints past a few turns. The Techniques page details exactly where each move stops.

Context engineering

What it names: Assembling what the model sees at inference time: retrieval, tool outputs, memory injections, window budget, compression.
Its strength: Agent pipelines, grounding, token economics — engineering the inputs so the model has the right material.
Where it lives in the session: One of the session's variables: part of the opening's five elements, and a maintenance concern as the window fills.
Seen from the session: Here is the direct answer to the question this page exists for — why is this not just context engineering renamed? Because context describes what the model is given; the session describes what the two of you do across time. A perfectly assembled context still drifts toward comfort if nobody steers; no retrieval stack reads the avant-texte, runs a trajectory audit, or decides that a concession was won by pressure rather than argument. Context is one variable among many — a session also runs on retakes, tolls, drift signals, external memory, closing, and operator vigilance. Inputs are engineered; conduct is steered.

Conversation design

What it names: Designing dialogue flows, personas and journeys — the chatbot UX tradition.
Its strength: Products and mass-facing assistants; scripted paths that must work for thousands of users.
Where it lives in the session: The design-time cousin of the opening frame.
Seen from the session: Design happens before; steering happens during. A design cannot pre-script the retake a live exchange will need — session engineering is the real-time practice that begins where the flowchart ends.

AI interaction design

What it names: The experience layer of human-AI interfaces: affordances, feedback, trust cues.
Its strength: Making capability legible and usable; shaping what interactions are possible at all.
Where it lives in the session: The room the session happens in — including whether the interface shows the model's reasoning.
Seen from the session: The interface shapes the possible; the method is what the operator does inside it. When the interface exposes the avant-texte, interaction design hands the operator this method's most valuable instrument.

Agent workflows

What it names: Decomposing a task into steps that agents execute, with defined handoffs.
Its strength: Automatable pipelines; scale; repeatability without a human at each step.
Where it lives in the session: Automation of stretches of a session — the parts that no longer need judgment each time.
Seen from the session: A workflow removes the human between steps; a session keeps the human as the control layer. The friction that validates a production has to come from somewhere — in a workflow it must be designed in, in a session it is supplied live.

Orchestration

What it names: Coordinating multiple models, tools and calls into a coherent system.
Its strength: Systems plumbing: routing, fallbacks, cost and latency management.
Where it lives in the session: The machinery underneath a session that uses several capabilities at once.
Seen from the session: Orchestration governs how machines cooperate; session engineering governs how a human and a machine think together. One is architecture, the other is conduct — a well-orchestrated system can still be badly steered.

Memory management

What it names: Deciding what persists: stores, injected memories, summaries, compression policies.
Its strength: Continuity across sessions; not paying twice for the same ground.
Where it lives in the session: The close's consolidation is memory management done by hand — export before it evaporates; context compression is memory management done to you.
Seen from the session: Memory is one variable, and the method's rule about it is blunt: a platform's history is not an archive. What matters survives because the operator externalized it, not because the system remembered.

Human-in-the-loop

What it names: Inserting human checkpoints into automated flows: review, approve, correct.
Its strength: High-stakes validation; catching what automation shouldn't decide alone.
Where it lives in the session: The minimal, sampled version of what the operator does continuously.
Seen from the session: HITL puts the human at gates; session engineering puts the human at the wheel. Sampled approval catches outputs; permanent steering shapes trajectories — including the drifts that no single checkpoint would flag.

Iterative prompting

What it names: Refining a result by re-asking: adjust, regenerate, converge.
Its strength: Converging on a shape when you can recognize it on sight.
Where it lives in the session: The raw gesture that the retake disciplines.
Seen from the session: Iteration without signals is correcting at random. The retake is iteration with a target: seven signals tell you when to intervene, three techniques tell you how — verbatim callback, short question, step-down request.

One mechanism worth naming: silent sampling. Faced with a long document, a model tends — by slope, not decision — to lean on the opening, the close and the salient passages, and to produce an analysis shaped like the whole without covering it. The counter-move lives on the method page: a coverage contract, three distributed quotes (start, middle, end), and a buried question you planted yourself. A claim of coverage is not coverage.

What next

The question this page settles returns in one line on the home page (FAQ), and the full method — signals, retake, avant-texte, protocol, test bench — is on the method page.