GPT-5 on context-budget rituals
ChatGPT-5 share link, 2026-04-09. I'd been losing context across sessions and wanted a framework, not another "use Projects" answer.
The prompt
I dumped four things: the v6 AGENTS.md draft, the mdfy Foundations bundle, the most recent 50 Cursor chat turns, and a one-paragraph statement of what I was trying to design (cross-AI context handoff). Then asked: "How would you structure the context you carry into a working session?"
What it built
A worksheet, not a heuristic. The structure was:
| Slot | Allocation | Source |
|---|---|---|
| Project context | 30% | AGENTS.md / CLAUDE.md / hub URL |
| Recent edits | 30% | Last N file diffs |
| Live conversation | 40% | This turn + last 4-6 |
The split is the part I actually use. Before this, I was loading "everything" — which meant the model got 12K tokens of static context and 1K of live discussion. The flip (live > context) sounds backwards but explains why "the AI keeps losing the thread mid-conversation" — the thread had no room.
The caveat
The worksheet over-fits to coding sessions. For research / writing, the ratios invert: live conversation shrinks (you're not iterating on a turn-by-turn loop) and project context expands (you're loading more reference material).
What I'm doing
- Default
?compacton hub URLs when pasted into Cursor — claws back 30% of the project-context slot for free. - Added a "Recent edits" Cmd+K command in the editor — pulls last 5 diffs into a paste-able block.
- For research, swap to
?full=1on bundle URLs — the whole bundle as context, no compaction.