Originally written in French. Translated by AI — the meaning has been preserved, not the prose.
Angle
What you file away at the end of a piece of work — the specification delivered, the minutes filed, the tender response sent — is not what will serve to do that work again. The document carries the result, never the reasoning that led to it, and holding a neutral layer of knowledge alongside it has always cost more than it returned. That is the calculation that has just flipped, and nothing else: the capital of a piece of work is no longer the document it produces.
Synthesis
Take an ordinary sequence. You write a specification for a development. Three weeks later, user stories are needed for the same feature. Then a pitch for the steering committee. Then a one-page note for the sales team. Then a tender response where the same feature has to be described to a buyer.
Five documents. A single piece of understanding work — the one that took place before the specification: the interviews, the constraints observed, the hypotheses ruled out, the reason it was settled that way.
That understanding is in none of the five. The specification carries the decision, not the deliberation. The pitch carries the promise, not the reservations. So with each document you start over: you reread, you ask again, you reconstitute. And the cost isn't visible on the first document, where it blends into the work itself. It appears on the second, where you redo what you have already done.
The reflex is to reuse the previous document — the pitch starts from the specification, the sales note starts from the pitch. Except that you then inherit what had been decided for another reader: a sentence written to reassure becomes a requirement, a slide formula becomes a development constraint. You think you are passing on knowledge; you are passing on presentation choices.
The solution has been known all along: hold a neutral layer in between — the sources kept, the ideas written for their own sake, with no output format. Nobody does it, and the reason isn't laziness. That layer speaks to no audience. Not to the customer, not to the developer, not to the committee. It can't be billed, it can't be presented, and someone has to keep it up to date while the deliverables have deadlines. The trade-off was rational: the cost was immediate, the gain deferred and uncertain.
That is the trade-off AI shifts, and nothing else. It doesn't make the layer more useful than before — it makes it affordable: writing a clean note, connecting it to what already exists, finding it again three months later now cost little enough for the calculation to flip. The conclusion therefore bears not on the tool but on what you file away at the end: what you keep after a deliverable is no longer the deliverable.
What doesn't change, on the other hand, is what the document is for. It never had value for what it preserved — it has value for the constraint it imposes: writing for a reader forces you to settle what you assert. You don't replace it with notes. You only stop asking it to stand in for memory.
Tensions / contradictions
The reasoning assumes the layer stays neutral. Yet what made it affordable — it is now produced from the sources, without your having to write it — is also what takes the author's hand off its form. The cost fell at the precise moment the guarantee of neutrality changed nature: it no longer rests on the discipline of whoever writes, but on the fidelity of whatever produces.
A second tension: the layer is worth what it makes findable, and a layer that fills up effortlessly also fills with what didn't deserve keeping. Nothing here says at what volume ease of production starts costing in quality of retrieval.
Questions
- How do you check that a layer produced from the sources is faithful to them, without rereading the sources?
- What, in that layer, must remain decided by someone rather than produced?