I audited 30 days of AI coding spend and found my biggest leak wasn’t the model

Last month I tracked every serious coding session with AI tools.

I expected the expensive part to be obvious: "big model = big bill."

Wrong.

The biggest leak was attention drift.

When my focus broke, prompts got messy, context windows bloated, retries stacked up, and one "quick fix" session became a 40-minute burn loop.

What I tracked

For 30 days, I logged:

• session start/end
• model used
• token/cost estimate per run
• number of retries
• whether I stayed in deep work or got pulled into feeds

The pattern was brutal:

• focused sessions: fewer retries, lower token spend, more shipped
• distracted sessions: 2-4x retries, higher token burn, less finished work

The 5 changes that actually worked

1. Scope before prompt (60 seconds)

   • write task, constraints, and done condition before touching AI.

2. Model tiering by task type

   • heavy reasoning only when needed
   • default to cheaper/faster models for scaffolding and cleanup.

3. Session cost visibility in real time

   • this changed behavior immediately.
   • I built TokenBar (https://tokenbar.site) as a $5 Mac menu bar tracker because I was tired of discovering cost damage at month-end.

4. No-feed coding blocks

   • I didn’t delete apps; I removed algorithmic feeds during build blocks.
   • I built Monk Mode (https://mac.monk-mode.lifestyle) for this exact problem ($15 one-time).

5. Hard stop rules

   • if retries > 3 without progress, stop, rewrite context, restart clean.

Counterintuitive takeaway

I used to think AI spend was purely a tooling problem.

It’s partly a focus problem.

Better attention made me write better prompts.
Better prompts reduced retries.
Fewer retries dropped cost.

Same tools, different behavior, very different outcomes.

If you’re a solo dev

Start with this daily 10-minute loop:

• review yesterday’s most expensive session
• identify if spend came from complexity or distraction
• set one no-feed build block for today
• cap retries per task

You don’t need perfect discipline.
You need better defaults.

If you’ve tracked your AI coding spend, what surprised you most?