I thought I had an AI model problem. I actually had a workflow problem.

For a while, I kept blaming model pricing.

“Claude is too expensive.”
“Codex burns too fast.”
“Cursor is cheaper but quality dips.”

Some of that is true. But it wasn’t the main leak.

The main leak was my workflow.

I’m a solo Mac dev building two products:

• TokenBar ($5) — a tiny menu bar app that shows real-time token and cost usage while I code: https://tokenbar.site
• Monk Mode ($15) — a feed-level distraction blocker for Mac, so I can keep useful app features but cut infinite-scroll traps: https://mac.monk-mode.lifestyle

After tracking my sessions for two weeks, I noticed a pattern:

• Focused sessions were consistently cheap.
• Distracted sessions were consistently expensive.

Not by 10%.
By 2–3x.

Where the money was actually leaking

It wasn’t giant prompts. It was tiny behavior loops:

1. “Let me quickly check X” during a run
2. Lose context
3. Come back and re-prompt from memory
4. Run one more variant “just to be safe”
5. Repeat

Each loop looked harmless.
Stacked together, it killed both spend and momentum.

What changed my output (and bill)

1) Session preflight (90 seconds)

Before each run, I now write:

• goal
• model
• stop condition
• max retries

This alone cut wasteful reruns hard.

2) 90-minute no-feed blocks

I don’t block entire apps. I block feed surfaces.
I can still use messages, docs, and search intentionally.
I just remove the infinite slot machine.

3) Live cost visibility while coding

Seeing spend in real time changed my behavior immediately.
I stopped “vibe rerunning” expensive tasks and got better at model routing.

4) End-of-session handoff note

I leave a 4-line note for future me:

• current state
• next exact step
• known risk
• first prompt to run next time

This prevents next-session context thrash.

The surprising part

When attention got cleaner, prompt quality improved.
When prompt quality improved, retries dropped.
When retries dropped, cost dropped.

So “focus” and “AI cost control” ended up being the same system.

If you’re feeling AI tool fatigue

Try this for 7 days:

• 1 preflight per session
• 1 no-feed block per day
• 1 live spend check during work
• 1 end-of-session handoff note

You don’t need perfect discipline.
You need fewer opportunities to drift.

That’s what finally made my nights productive instead of just “busy.”