Prompting Astra: deleting old rules beats adding new ones
The author argues Astra is unusually sensitive to stale rules in skills and AGENTS.md, that constraints written for weaker models overconstrain it and burn context, and that input above 272K tokens is repriced at 2x.
YOU'RE PAYING GPT-6 ASTRA $50/M OUTPUT TO OBEY RULES WRITTEN FOR A DUMBER F*CKING MODEL Fewer rules instead of a longer prompt The trick is not a bigger instruction file It is deleting what Astra no longer needs Most teams keep stacking the same things another rule another warning another "always do this" another checklist another skill another patch for a weaker model That is how the bill gets fucking expensive Astra is unusually sensitive to what hides inside skills and AGENTS.md Old scaffolding written for weaker models can overconstrain it and burn context Same model Same repo Completely different behavior It may delegate less than your workflow expects unless you say so explicitly It may test more broadly than a small task needs And above 272K input tokens the whole request is repriced at 2x input and 1.5x output Your instructions do not need to disappear They only need to stop fighting the model The old prompt gets the habit The new model gets the room Full breakdown below ↓
This post quotes another post: x.com/fluixoo/status/2106287658053840899