Idea

A cap on a tool's instructions separates an embedded core from an extension loaded on demand

Info

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

Main idea

The instructions of a ChatGPT Project are capped at roughly 8,000 characters. A rule set that exceeds that cap is not summarized: it is cut into two levels of a different nature. A core fits inside the instructions and holds what the engine cannot start without — what a context is, where it is stored, which commands exist, the file structure, the invariants, and the address of the rest. An extended instruction lives elsewhere, in a document the assistant knows how to open, and carries the detailed method: ingestion criteria, reading strategies, edge case handling.

The core is not merely short: it knows where to find the extension and orders it loaded at the start of a conversation. What holds the whole together is therefore not the brevity of the core but the pointer it contains.

The cut does not follow the importance of the rules — the extension is often richer than the core. It follows the moment they become useful: what must be true before knowing anything about the work at hand, and what only serves once you know what you are doing.

Why it matters

It turns an imposed limit into a design principle. Without the cap, you would have stacked the rules into a single block and discovered later that the assistant handles poorly whatever is buried in the middle of a long instruction. The cap forces a split that would have been the right answer anyway.

And it gives a sorting rule that applies elsewhere: faced with a rule set too large to fit, look for what must be true at startup, and leave the rest callable.

Nuances and limits

The split has a cost: it adds a dependency. If the extension is not reachable — document moved, access revoked — the engine starts with a core that believes it knows where to go. A single block would not have that failure mode.

And the boundary between the two levels is not stable: a rule judged to be detail becomes fundamental as soon as it is broken often enough, and nothing signals that it should be moved up into the core.

Open questions

  • What indicates that a rule lodged in the extension should have been in the core, given that its absence at startup produces no error but a slow drift?