Practice
Context Does Not Fit in a Prompt
What separates a plausible legal analysis from an enforceable deliverable lies almost entirely in what is not written in the document.
Take a shareholders’ agreement review, and follow it for real, step by step, as it unfolds in a firm. The draft arrives from the opposing side on a Thursday evening. The partner opens it, and within fifteen seconds, before even reading a clause, they already know three things: what the balance of power is in this negotiation, what the client’s two or three red lines are, and what drafting register the opposing firm uses when it wants to hide a concession in a definition.
None of this is in the document. Everything is in the layer above: the client’s history, the three comparable transactions handled last year, the internal note from six months ago where the team settled a warranty question. The partner does not read a contract. They read a contract through a memory. And it is this memory, not their knowledge of corporate law, that makes the difference between their reading and that of an equally competent lawyer who is a stranger to the matter.
One must dwell on these fifteen seconds, because they contain the essence of what we are trying to understand. What the partner mobilizes in an instant is written nowhere in usable form. These are not pieces of knowledge one could find in a manual or a precedent; they are accumulated judgments, inferences drawn from dozens of prior matters, a familiarity with a client and an opponent that is recorded in no file. This knowledge is real, it is decisive, and it is tacit. It is precisely this tacit part, invisible in the document, that separates an expert reading from a competent but blind one.
The document is the visible object. The work happens in the invisible layer.
What a copilot cannot produce
Now ask a copilot to analyze the warranty clause of that same agreement. The result will often be correct, sometimes remarkable. It will identify the caps, the baskets, the durations, the exclusions, and comment on them with real technical accuracy. But it will have done so out of matter: without knowing that this client has never accepted a cap below a certain threshold, without knowing that the firm has drafted this warranty in a particular way for years, without knowing what was conceded on the previous transaction in exchange for what.
It is the exact equivalent of entrusting the review to a brilliant associate given neither the matter, nor the client, nor the firm’s acceptability criteria. They will return something plausible. They will not return something operational. And the difference between the two is precisely the craft. A plausible deliverable must be entirely rechecked before use, because one does not know whether it integrates what matters; an operational deliverable can be committed, because it was produced with knowledge of the matter.
An out-of-matter answer is a plausible answer. It is not a deliverable.
This distinction is not a theoretical refinement. It decides the time actually saved. An opinion produced out of context forces the senior to mentally reconstruct everything the AI ignored, then to check that nothing essential was omitted. The apparent gain at production is taken back at review. The work was not accelerated, it was displaced, and often displaced toward the most expensive person in the chain, the very one whose time was to be preserved.
The objection: “just put the context in the prompt”
This is the reflex answer, and it deserves to be taken seriously, because it is right in theory and wrong in practice. Yes, one can paste into a prompt the client’s history, the methodology, the prior trade-offs. But do the math of what this really supposes. One would have to, at each interaction, manually gather the right context from among hundreds of matters, format it, prioritize it, reinject it, check that it is current, and start over at the next session because the model will have kept none of it.
And that is only the optimistic version. In a real firm, the right context is not known in advance: knowing which precedents are relevant to this clause is already a legal act, the very one we hoped to delegate. To ask the user to provide the context is to ask them to have already done the work the tool was supposed to do. We go in circles, and the circle closes in the worst place: on the expert whose burden we wanted to lighten.
Context does not fit in a prompt for the same reason a library does not fit in a sentence: it is not a question of size, it is a question of structure, persistence and governance. A prompt is an event, one-off and with no tomorrow. A matter is a story, continuous and accumulative. You do not fit a story into an event, and enlarging the event changes nothing, because what is missing is not space, it is permanence.
Putting the context in the prompt is asking the user to redo by hand the work of the layer.
What tacit context really covers
To understand why this context resists the prompt, one must look at what it is made of, for it is vaster and more diverse than one imagines. There is first the memory of matters: what was handled, decided, conceded, and why. There is then knowledge of the client: their preferences, their thresholds, their risk tolerance, their negotiation history. There is the firm’s methodology: the way, here and nowhere else, one drafts such a clause, structures such an opinion, prioritizes such a requirement. There is finally the reading of the opponent: what such a firm does when it uses such phrasing. Each of these registers is a layer of context, and none is written in the document being analyzed.
What strikes, when one enumerates these registers, is that none reduces to information one could simply retrieve and paste. They are interpretive knowledge: they say not only what happened, but how to read it. Knowing that a client refused a cap last year has value only when tied to the fact that they refused it in a precise balance of power, for a precise reason, that holds or no longer holds today. This context is not a datum, it is a reading, and a reading is not pasted into a prompt: it is built, maintained, governed over time.
This is why the context layer cannot be improvised at each interaction. It must precede the question, be structured, kept current, made available at the right moment without anyone having to gather it. It is not an accessory added to a capable model; it is the infrastructure without which the model’s performance stays groundless. A brilliant model fed poor context returns a poorly situated analysis; an ordinary model fed rich context returns a deliverable that holds. The difference plays out not in the model, but in what surrounds it.
Legal context is not a datum you paste. It is a reading that is built and maintained.
Why this layer is the asset, and the model the component
The industry spent two years improving the quality of isolated analysis, and that was useful: it proved the models kept the linguistic promise. But legal work does not stop at isolated analysis, it begins there. Everything that plays out between the model’s answer and the firm’s deliverable, the insertion into a matter that has a history, the validation according to a methodology that belongs to the firm, the articulation in a workflow, the traceability that makes the AI enforceable, exists nowhere in the models.
One must see that this context layer is not an improvement of the model, it is a thing of another nature, placed beside and above it. The model brings linguistic capability, general competence, processing power; the layer brings what the model does not contain and never will, because it belongs not to language in general but to this firm in particular: its memory, its trade-offs, its criteria, its history with this client. No improvement of the model produces this part, because it is not in the model to be improved.
And it is strategically good news, for that is where the defensible value lies. Models will keep improving and commoditizing; their raw performance will align. What will remain rare, proper to each organization and hard to replicate, is precisely this layer of context, memory and methodology placed above them. It is this that MAX builds: not to produce one more answer, but to hold what the models do not hold. The model is the component. The layer is the asset.
The model produces an analysis. The layer produces a deliverable. Between the two lies the whole craft.
