I used to end most days with two bad surprises:
1) "Why is my AI bill this high?"
2) "How did I spend 2 hours on X/YouTube when I said I was shipping?"
After a few painful weeks, I stopped treating these as separate problems.
They were both leak problems.
Here are 9 changes that actually worked for me.
1) I started tracking cost per session, not monthly totals
Monthly numbers are too late. By then, the damage is done.
I now check token/cost per work session and ask: "Did this session ship anything?"
2) I separate "explore" sessions from "ship" sessions
Exploration is where token burn explodes.
Shipping sessions are tighter and cheaper.
Different intent = different model settings + different time box.
3) I downgrade models by default
I used to default to the most expensive model for everything.
Now expensive models are opt-in only when quality truly matters.
This one change cut a surprising chunk of waste.
4) I kill retry loops fast
When I see myself prompting the same thing 5 different ways, I stop.
Usually that means bad context or unclear spec, not "model issue."
5) I block feeds during build windows
Not websites — feeds specifically.
Search/docs are useful. Infinite feeds are where time disappears.
6) I run short, aggressive focus sprints
45–60 min build sprint.
10 min reset.
Repeat.
If I keep feeds available "just for a sec," the sprint dies.
7) I score days on two numbers
- dollars spent on AI
- deep-work hours protected
When both improve together, shipping gets easier.
8) I built tools for my own pain points
I made:
- TokenBar: a tiny Mac menu bar app to watch live AI token spend while I work
- Monk Mode: a Mac app that blocks distracting feeds so I can stay in build mode
Honestly, both came from me being tired of my own bad habits.
9) I review "expensive + unshipped" sessions weekly
Any session that cost real money and shipped nothing gets reviewed.
Not for guilt — for pattern detection.
That review is where most improvements came from.
The key lesson for me:
attention spend and token spend are coupled.
When attention is scattered, prompts get sloppy and iteration gets expensive.
When focus is clean, both output and cost improve.
If you're building with AI daily, what was your biggest hidden cost leak?
Top comments (0)