I Tracked Every AI Token and Every Doomscroll Trigger for 7 Days (Here’s What Actually Worked)

Last week I had two problems at the same time:

1. My AI coding sessions felt expensive in random bursts.
2. My focus kept getting shredded by “just 2 minutes” of scrolling.

So I ran a simple 7-day experiment on my Mac:

• Track token/cost behavior while coding
• Track feed/open-tab behavior while working
• Change one thing at a time

Here’s what actually moved the needle.

The setup

I’m a solo builder, so context switching is brutal.

My stack was basic:

• AI coding sessions throughout the day (Claude/Cursor/Codex style workflows)
• A menu bar cost tracker for token visibility (TokenBar)
• A feed-level blocker during work blocks (Monk Mode)

Nothing fancy. Just visibility + constraints.

Day 1-2: The painful baseline

What I noticed immediately:

• Most token burn didn’t come from “hard coding,” it came from retries after vague prompts.
• The most expensive sessions happened after I had already broken focus and came back scattered.
• Screen time spikes happened in tiny 30-90 second checks that turned into 12-20 minute detours.

The big insight: token waste and doomscrolling were linked.

When I was focused, prompts were sharper and cheaper.
When I was distracted, I paid twice: once in time, once in tokens.

Day 3-4: One rule before every AI run

Before hitting Enter, I forced this mini checklist:

• Goal in one sentence
• Budget cap (soft max)
• Stop condition (what “done” means)

That alone reduced reruns a lot.

I also kept live token/cost visible in the menu bar so I stopped “flying blind.”

Result: less emotional guessing, more deliberate decisions.

Day 5-6: Feed-level blocking during deep work

I didn’t block the entire internet.
I only blocked infinite feeds during maker blocks.

That was the difference.

Because the blocker acted at feed-level, I could still access useful pages/docs, but not get dragged into algorithm loops.

This made starting work easier and, more importantly, made staying in flow possible.

Day 7: What changed

By the end of the week:

• Fewer runaway AI sessions
• Cleaner prompts
• More completed tasks per day
• Less mental fatigue at night

The important part isn’t “discipline.”
It’s reducing invisible friction and invisible leaks.

The 5 practical takeaways

1. Track what hurts in real time
   If you can’t see cost/focus drift while it happens, you’ll rationalize it later.

2. Never start an AI session from a fuzzy brief
   Vague in = expensive out.

3. Use budget + stop conditions
   Treat each run like a bounded task, not a wandering conversation.

4. Block feeds, not everything
   Overblocking backfires. Precision blocking sticks.

5. Assume distraction multiplies token spend
   Focus quality directly affects AI spend quality.

Tools I used

• TokenBar — a simple Mac menu bar tracker for real-time token/cost visibility.
• Monk Mode — a Mac distraction blocker that can target feed-level doomscroll traps.

You can absolutely replicate the framework with any tools, but the key is this combo:

Visibility + constraints + environment design.

That’s what made this week work.

---

If you’re building with AI daily, don’t only optimize prompts.
Optimize your attention system too.

That’s where the real savings are.