I had a weird month:
- My AI coding output was up
- My bank account was down
- My focus was gone by 3PM every day
I finally noticed the pattern: I was leaking in two places at once — token spend and attention spend.
So I treated both like one system.
Here are the 7 rules that helped me ship more with less chaos.
1) Track token spend in real time, not at end of month
If you only check provider invoices, you’re already too late.
I started running TokenBar in my macOS menu bar so every prompt had a visible cost attached to it.
That one change killed “just one more retry” behavior.
2) End every coding block with a 5-line handoff note
Most expensive token waste comes from re-loading context after breaks.
I now end each session with:
- what I changed
- what’s broken
- next exact step
- files touched
- constraints
Next session starts faster, with fewer re-reads.
3) Cap retries per task
I use a simple rule: max 2 blind retries.
If the model misses twice, I rewrite instructions or scope down.
This dropped pointless token churn hard.
4) Remove feed-level distractions, not just websites
Website blockers helped, but I still got trapped inside “allowed” sites via recommendations.
I switched to Monk Mode and blocked feeds directly.
Result: I can still use platforms when needed, without getting pulled into algorithm loops.
5) Batch model-heavy work
Instead of context-switching all day, I batch:
- architecture prompts
- code generation
- test/debug loops
Fewer context resets = lower token burn.
6) Protect the first 90 minutes of the day
No notifications, no feeds, no “quick checks.”
Deep work first. Reactive work later.
This alone made my outputs feel 2x cleaner.
7) Review both budgets weekly: dollars + attention
Every Friday I check:
- AI spend by workflow
- hours lost to distraction loops
- what shipped vs what I felt busy doing
Brutal but useful.
What changed after 3 weeks
- Fewer giant token spikes
- More shipped PRs per day
- Less doomscroll residue in my head
- Better evening energy
The big lesson: cost control and focus control are the same game for AI builders.
If you optimize only one, the other will quietly wreck it.
Top comments (0)