For a while I thought I had two separate problems:
- My AI coding spend kept creeping up every week.
- My nights kept disappearing into feeds, tabs, and “just one more scroll.”
Turns out it was one system failure.
When my attention was fragmented, my prompts were worse.
Worse prompts meant more retries.
More retries meant a bigger bill.
The pattern I kept seeing
On focused nights, I’d ship more and spend less.
On scattered nights, I’d “work” for hours and burn tokens.
The biggest cost spikes were rarely from hard engineering.
They came from:
- re-running prompts with fuzzy scope
- context switching between coding + feeds
- using heavyweight models for lightweight tasks
The nightly shutdown ritual that fixed it
I now do this every night before I stop:
1) Review the day’s expensive runs (5 min)
I check which sessions actually burned the most and write one line for each: “why was this expensive?”
2) Set tomorrow’s model routing (3 min)
I pre-decide what gets premium reasoning and what stays on cheaper/default models.
3) Block feed surfaces for the next deep-work block (2 min)
Not deleting apps. Just removing the algorithmic feed loops that hijack context.
4) Write a 3-bullet kickoff brief (5 min)
So I start with clear prompts, not morning chaos.
Tools I built for this
I ended up building two tiny Mac apps because I couldn’t find exactly this combo:
- TokenBar ($5, one-time): live token/cost visibility in the menu bar while I work.
- Monk Mode ($15, one-time): feed-level blocking so I keep utility, lose the infinite scroll trap.
I made them for myself first, but they became the backbone of this routine.
Results after 2 weeks
- Fewer “mystery” cost spikes
- Less retry churn on prompts
- Cleaner deep-work sessions at night
- More shipped work with less mental drag
No productivity magic.
Just better feedback loops and fewer attention leaks.
If your AI spend feels random, check your focus system before you blame model pricing.
Top comments (0)