For a while my nights looked productive:
- 4+ hours in Claude/Codex
- tons of prompts
- tabs everywhere
- almost nothing actually shipped
The expensive part wasn’t the model.
It was context-switching, retries, and doomscroll breaks between runs.
I started using a tiny nightly system:
1) 2-minute preflight before every heavy run
I write a 4-line brief: goal, files, done condition, max spend.
That alone killed a lot of vague reruns.
2) Live cost visibility while I’m coding
I built TokenBar (https://tokenbar.site) as a $5 Mac menu bar tracker so I can see token/cost drift in real time instead of discovering it on the bill.
3) Remove feed traps, not the whole app
I built Monk Mode (https://mac.monk-mode.lifestyle) as a $15 Mac app that blocks algorithmic feeds while keeping useful surfaces.
That stopped the “quick scroll → lost context → expensive retry” loop.
4) Hard stop rule
After 2 failed attempts on the same task, I stop and rewrite the brief.
No third blind retry.
What changed
- fewer panic retries
- cleaner prompts
- lower AI spend per shipped task
- longer focused blocks at night
If you’re a solo dev, this combo matters more than chasing every new model release.
Boring systems beat hype.
Top comments (0)