Main idea

A workspace fed for weeks — dozens of sources, notes, session logs, successive states — cannot be put back whole into an assistant session. So the choice is not between loading everything and loading nothing: it is deciding, before the session, what always goes in and what only goes in on request.

The core is what always goes in, and it is short: the mission (why we are here), the current state (where the work stands), the plan, the decisions made, the contradictions still open, a resumption sheet. A few files are enough, and they make the assistant operational immediately without giving it noise. The rest — raw sources, notes, logs — stays on disk, accessible but absent, and is called only if the session needs it: a specific source to analyze, a note to enrich, an older state to compare against.

This distribution does not hold unless it is written down somewhere. A loading manifest states the order and the priority: what always loads, what is recommended depending on the phase under way, what stays available without being loaded. It is a surgeon's tray — the essential instruments are already laid out, the others are within reach, and nothing useless takes up the table.

Layer added by "Running my Contexts directly inside ChatGPT" (2026-09-20). The rule also holds when the workspace is not a disk but a Google Sheets workbook, one tab per type of information. On opening a Context, the assistant loads the mission, the instructions, the state, the synthesis, the tasks, the decisions, the contradictions and the handoff sheet; the raw sources, the atomic knowledge and the older deliverables stay in the workbook and are read only when a question calls for them. Changing the medium does not change the distribution, which indicates that it follows from the model's attention window and not from the nature of the storage.

Why it matters

This makes loading a design decision, taken once and revisable, rather than a trade-off redone at every session opening by someone who has other things to do. Sorting left to the moment of the session is done badly: you load what is at hand, not what matters.

It also gives a criterion for judging a workspace: if no short subset exists that allows the work to be resumed, then the current state has never been written down anywhere.

Nuances and limits

A human expert doesn't need this setup: they have internalized the essentials of their file and go looking for a specific document when one is missing. An assistant cannot internalize — it has only what it is given, every time — and it is that difference that makes explicit selection necessary here and superfluous there.

And the core goes stale: maintained by nobody, the current state describes a situation from three weeks ago, and it is loaded all the more faithfully for being a priority.

Open questions

• What signals that a file kept outside the core should have been in it, given that its absence produces no visible error?