DEV Community

Henry Godnick
Henry Godnick

Posted on

I was burning money and time every night building AI apps — here’s the workflow that fixed both

#productivity

If you’re building with AI every day, you already know the trap:

  • You start with “quick iteration”
  • You open one social feed “for 2 minutes”
  • You context-switch 40 times
  • You rerun prompts because your brain is scattered
  • You end the night with less shipped and a bigger AI bill

I hit this wall hard while building two Mac apps solo.

What I was doing wrong

For weeks, my workflow looked “productive” but wasn’t:

  1. Keep Claude/Codex/Cursor open all day
  2. Jump into feeds whenever a build runs
  3. Return with broken context
  4. Re-prompt from scratch
  5. Burn tokens fixing preventable mistakes

I wasn’t losing to model quality. I was losing to my own loop.

The 5 changes that fixed it

1) I started tracking token spend in real time

Seeing cost live changed my behavior instantly.

When I could watch spend climb while I kept “just trying one more prompt,” I got disciplined fast.

I built this into my menu bar app TokenBar (tokenbar.site) so I can’t ignore it.

2) I added a preflight before every AI run

Before hitting enter, I force myself to write:

  • exact objective
  • files in scope
  • definition of done
  • max retries

This reduced garbage prompts and expensive reruns.

3) I blocked feed-level distractions during build sprints

Not just app blocking. Feed-level blocking.

Because opening YouTube for “one tutorial” and landing on the home feed kills the session.

I use my app Monk Mode (mac.monk-mode.lifestyle) for this.

4) I switched to shorter execution loops

Instead of giant prompts and giant context windows, I work in tighter chunks:

  • plan
  • execute
  • verify
  • stop

Smaller loops = fewer drift errors = lower cost.

5) I added hard stop conditions

Every task gets one:

  • time cap
  • token cap
  • retry cap

No cap = emotional decision-making = budget leaks.

My current nightly stack

  • Cost visibility: TokenBar
  • Focus protection: Monk Mode
  • Execution: Claude/Codex/Cursor in short loops

This combo made my nights way less chaotic.

Results so far

  • fewer pointless reruns
  • lower token waste
  • less doomscrolling between runs
  • more real shipping momentum

Still improving, but this is the first setup that consistently works.

If you’re building solo with AI, treat focus and cost tracking as core infrastructure, not “nice to have.”

Top comments (0)