I used to treat two problems as unrelated:
- My AI token bill was climbing every week.
- My screen-time graph looked like a meme.
Turns out they were the same bug: low-friction context switching.
Here are 7 habits that finally worked for me.
1) Put a live cost meter where you can’t ignore it
When usage is hidden, prompts get lazy.
I started keeping TokenBar open in my Mac menu bar so every model run had a visible price tag. The moment cost became visible, I stopped dumping giant “do everything” prompts and started writing scoped asks.
2) Use a “3-line brief” before every expensive run
Before I hit enter, I force myself to write:
- Outcome
- Constraints
- Definition of done
This one change cut retries hard.
3) Cap every work block to 20–30 minutes
Unbounded sessions are where cost explodes. I timebox each agent run, then re-scope.
Short blocks = fewer wandering prompts.
4) Kill the feed during build windows
I used Monk Mode to block social feeds at the feed level while still letting me access specific tools/messages when needed.
Not seeing endless feeds means fewer “quick checks” that become 25-minute detours.
5) Add a “retry tax” rule
If a prompt fails twice, I rewrite from scratch instead of nudging forever.
Most runaway cost came from emotional retries, not hard problems.
6) Track cost + attention in one daily note
I write down two numbers at night:
- AI spend today
- Doomscroll minutes today
If one spikes, the other usually did too.
7) Build a boring fallback plan for outages
When Claude/Cursor/Codex hiccup, panic-clicking can burn serious budget.
Now I keep a fallback playbook:
- smallest possible task
- single model
- hard stop after 20 minutes
No drama, less spend.
What changed after one week
- Lower token spend
- Fewer context resets
- Better output per prompt
- Less late-night “I was just checking one thing” scrolling
If you’re a solo builder, don’t optimize prompts in isolation. Optimize attention + prompt quality together.
That was the unlock for me.
Top comments (0)