I used to treat these as two separate problems:
- My AI bill kept climbing
- My focus kept collapsing at night
Turns out they were the same leak.
When I doomscrolled between coding sessions, I’d come back with broken context, re-prompt everything, and burn way more tokens than needed.
Last week I forced myself into a simple system. Nothing fancy. Just habits I could actually keep.
Here’s what worked.
1) Track cost during the session, not at month-end
Month-end billing dashboards are too late. By then, the bad habits are already reinforced.
I started watching token spend live in my menu bar while coding. That tiny feedback loop changed behavior instantly:
- I stopped sending giant “do everything” prompts
- I split tasks into smaller passes
- I quit retry-spamming when the first answer was messy
I use TokenBar for this (Mac menu bar token/cost tracker). It sounds small, but seeing cost in real time made me code like someone paying cash.
2) Define a “ship unit” before opening AI tools
Before Claude/Cursor/Codex, I now write one sentence:
“In this session, I’m shipping X.”
If a prompt doesn’t move that ship unit forward, I don’t send it.
This killed a lot of expensive curiosity prompting.
3) Run 45-minute deep-work blocks with zero feed access
My worst token spikes happened after “quick breaks” that became 25-minute feed rabbit holes.
Now I block social feeds during build windows and only allow them in explicit breaks.
I use Monk Mode (Mac distraction blocker with feed-level blocking) so I can still use tools I need without getting sucked into infinite scroll.
The weird result: fewer prompts, better prompts, faster merges.
4) Use a 3-pass prompting pattern
Instead of one giant prompt, I do:
- Pass 1: architecture/options
- Pass 2: implement one narrow slice
- Pass 3: tighten/tests/refactor
This keeps context cleaner and reduces token waste from repeated broad re-reads.
5) Cap “investigation depth” early
I used to ask AI to inspect half the repo “just in case.”
Now I cap investigation:
- max files first pass: 3–5
- if still unclear, expand intentionally
Most bugs don’t need a repo-wide tour on the first attempt.
6) Add a mid-session reset trigger
When I notice myself tab-hopping or re-reading the same output, I run a 2-minute reset:
- stand up
- re-state the ship unit
- check current token spend
- choose the next smallest step
This is boring and works absurdly well.
7) End each session with one line of review
At session end, I log:
- shipped outcome
- rough spend
- one waste pattern I noticed
This made the next session cleaner every time.
7-day result (for me)
- lower token waste from “prompt panic” and context churn
- less doomscrolling during prime build hours
- more finished units per night
Not magic. Just visible feedback + feed control.
If you’re building solo and feel like cost + focus are both slipping, try fixing them as one system, not two separate problems.
That shift made the biggest difference for me.
Top comments (0)