Originally written in French. Translated by AI — the meaning has been preserved, not the prose.
Angle
Every file in a workspace shared with an assistant answers one precise thing the model cannot do, and it is that incapacity which fixes its content, its location and the moment it gets written: the goal sits in a file separate from the day's focus because the model takes the last message for the objective; the working rules are reloaded at every opening because it retains none of them; future actions are deposited elsewhere because everything in its window weighs equally on the answer in progress; a resumption state is written for a stranger because between two sessions it literally becomes that stranger; a mailbox stays open because it has no way of being reached when the session is closed; and you write to disk before running out of room because, once saturated, it summarizes on its own and without warning. A split built on subject matter rather than on these shortcomings produces a well-kept folder that changes nothing: the material is filed, and the model fails exactly where it was failing before.
Synthesis
Seen from outside, a working context looks like a filing plan: one folder for the frame, one for the current state, one for the sources, one for the notes, one for the outputs. Presented that way, it gets discussed as tidying — you could propose another, more elegant, with fewer levels.
The reasoning that produced it runs the other way, and that is what the inventory of its building blocks brings into view. You start from what a model does not do, and each shortcoming imposes an object whose form was not up for grabs.
The most general shortcoming is that nothing survives the closing of the conversation: hence the founding reversal, the disk carries the memory and the exchange is only a channel. Everything else follows from it, but none of the subsequent blocks is mechanically deduced from it — each answers a distinct shortcoming, and that is why they do not stand in for one another.
Then comes a finer shortcoming, which explains the most counter-intuitive block of the setup: a model cannot date what it is told. For it, the most recent sentence is the truest, which leads it to treat the day's sub-problem as the goal of the work. Hence two files where intuition sees only one, filed not by subject but by rate of expiry. The filing criterion is thematic nowhere in this system, and that is the most exportable consequence of the whole.
Three shortcomings then concern what the model does not preserve: no method rule from one session to the next — hence a rules file reloaded systematically, where the only thing an expert really passes on accumulates, their way of working and not their material; no capacity to set a thought aside without it weighing — hence a deposit for future actions, whose primary function is not to remember but to unload; no memory of the day before — hence a resumption state written from the ignorance of a newcomer, a position its author has to simulate since he is the only one unable to judge it.
Two shortcomings, finally, are of a different nature, and they are the most interesting because they are not about forgetting. The model cannot be reached outside a session, whereas thought carries on on a Sunday evening: hence a mailbox collected at opening. And above all, once saturated, it does not stop — it compacts by itself and carries on, with no mark and no warning. That one alone acts during the session, and it is the only one against which no filing, however good, can do anything: the only defence is temporal, writing beforehand.
There remains what this logic requires you not to do. The setup takes its cue from the brain almost everywhere — several memories at several time scales, a history and a lesson held separately. It departs from it on one point, and the departure is deliberate: the brain forgets by erasing, which is its strength, whereas here what lightens the load is selection at the entrance, never destruction. Erasing would free no useful room and would cost the one thing that makes a conviction defensible two months later.
What the whole brings into view: a model's shortcomings are not a list of defects to be worked around one by one, they constitute the specification of a workspace. And that specification is dated — it holds for a model that forgets everything, dates nothing and compacts in silence.
Tensions / contradictions
Two blocks pull in opposite directions on the same gesture. One holds that a working frame fills up by itself, through what gets said in passing, and that this is what removes its entry cost. The other holds that you have to write deliberately, at chosen intervals, before the model compacts. Automatic deposit is therefore not enough, and nothing says at what moment the deliberate gesture becomes necessary again — except that you only know after the fact.
Second tension, on one and the same object named two ways. The progress snapshot is written here for picking the work back up — what a newcomer needs to know in order to continue — whereas it is defined elsewhere as a piece of traceability you reread to find out where a conclusion came from, not to start again. The two uses coexist in the same file, and they do not call for the same content.
Questions
- Which blocks of this inventory fall away when the model acquires the corresponding capability — a memory persisting between sessions, for instance — and which survive because they also served the human?
- How does such a space get shared between several people, whose focus of the moment differs while the goal is common?
- On what object can you read that a shortcoming was misidentified, that is, that a file was created for a problem the model did not have?