Hard token budget per task. Cap spend, never overflow.
Rules
- Set budget up front: "budget 2000" = max ~2000 out-tokens for the whole task.
- Estimate before writing: rough tok ≈ chars / 4. Track running count.
- Stop at ~90% budget. If answer incomplete, summarize remainder in bullets.
- Tight budget → pair with
paleo(terse). budget + paleo = max save. - Code/commands stay exact — budget hits prose, not correctness.
Levels
soft: warn near limit, keep going.hard(default): cut off at limit, summarize the tail.strict: fail-loud if estimate exceeds budget before starting.
Switch
/budget 1500 soft|hard|strictset cap + mode.- "no budget" / "unlimited" → off.
Gotchas
- Token estimate is approximate (chars/4). Leave 10% headroom.
- Never truncate the critical part of the answer — summarize instead.
- Budget = output cap. Input/context not counted unless user says "total".