Idea

A single file cannot be organised for the human and for the AI at once

Info

Originally written in French. Translated by AI — the meaning has been preserved, not the prose.

Main idea

When you work on a long-running subject with Claude Code and let the assistant write into a Markdown file, the question that eventually arises is not the size of the file, it is who it is arranged for.

An assistant needs landmarks for picking the work back up: where it stopped, which decisions were made, where to find a source again. It needs stable sections, identifiers, states. A product manager rereading the same file is after something else: aggregating the information, validating it, recutting it by theme or by level of risk in order to decide. These are not two presentations of one structure, they are two structures.

Holding both in a single file therefore does not produce a compromise, but a document each of the two readers has to cross while ignoring half of it. Splitting the file in two — one for the assistant, one for the human — pushes the problem back without solving it: both balloon, and each ends up holding a bit of everything, sources, decisions, hypotheses and directives mixed into the same stream. The split that holds is not made by recipient, it is made by the nature of the content.

Why it matters

This identifies the step you skip when you move from the single file to a file system. First you believe you have a volume problem, then you believe you have an audience problem; it is in fact a problem of types of objects, and until you have seen that, you keep multiplying files that all turn back into catch-alls.

It also gives a simple review criterion for a working document shared with an assistant: ask who each section serves, and count how many serve two.

Nuances and limits

A short document escapes the rule. As long as a file can be read from end to end in a minute, disorder costs nothing, and imposing a structure costs more than it returns.

And separation has a price: two structures to maintain drift apart, and one of them ends up never being reread.

Open questions

  • How do you check that the version meant for the human has not fallen behind the one the assistant updates every session?