Mini-Series Part 2: The "Token Crash" & The Keyboard Epiphany

In Part 1, I was riding high. I had Chrome, Edge, and Firefox all running simultaneously, tripling my speed on a massive D365 rebilling project. I felt like I’d cracked the code to being "efficiently lazy."

Link: Mini-Series Part 1: The "Manual Hell" Rebill Project · Clief Notes

But the victory was short-lived. I was driving the car at 100mph, but I didn't realize how much fuel I was burning.

The Crash: 3 Hours to Zero

The next morning, I fired up all cylinders. It was beautiful—until it wasn't. In under 3 hours, I burned through my entire token allotment. I was dead in the water with a 1.5-hour wait until my limits reset. With the clock ticking and sales reps still needing updates, I had to go back to doing it manually.

And that’s when the "Aha!" moment happened.

The "Tab" Breakthrough

While I was clicking and typing manually to pass the time, I realized something: I wasn't really using the mouse. I was mostly hitting Tab and Shift-Tab to navigate the D365 fields.

I asked Claude: "Can Playwright just 'Tab' through the browser instead of searching for elements and taking screenshots for every move?"

Claude’s answer: YES. My jaw dropped.

From "Visual Driving" to "Macro Speed"

The old workflow relied on Claude "seeing" the screen via screenshots to make sure it was in the right spot. It was accurate, but heavy. I had Claude rewrite the interior workflow to a deterministic loop:

The Setup: I open the URL, set the filters, and select the first row.
The Loop: * Tab to Invoice → Enter.
Repeat.

I added small buffers (100ms between tabs, 4s for page loads) to ensure D365 could keep up.

See the video for how the flow runs in real life!

The Result: 90% Savings & 10x Efficiency

I asked Claude to evaluate the two methods. The numbers were staggering:

Old "Visual" Flow:

Tool Calls: ~500 calls per 50 orders
Token Usage: 100% (High Burn)
Logic: Screenshot-dependent (Slow)

New "Tab" Flow:

Tool Calls: 1–4 calls per 50 orders
Token Usage: 10% (90% Reduction!)
Logic: Command-driven (10x Faster)

By stopping the constant screenshots and "thinking" cycles, the workflow became a lean, mean, rebilling machine.

The "Accounting Brain" Safeguard

The new flow was fast—too fast. After about 5 loops, it would occasionally lose its place and start "going rogue," tabbing and entering into the wrong fields.

I needed a safeguard. I told Claude: "The Sales Order is always a 10-digit number. The Customer Account next to it is always 9 digits."

I had Claude add a check after every "Arrow Down":

The Check: Read the current field. Is it 10 digits?
Yes: Continue.
No: STOP.

If it got off track, the script would halt. I’d manually click the right column, hit "Go," and it would fly again. It wasn't worth building a complex "self-correcting" logic for a one-time assignment; I just needed a "dead man's switch."

The Finish Line

I had one rep with 100 orders. I split his customers across my three browsers. In 30 minutes, he was done. In that same 30 minutes, I was actually doing my other accounting work.

What I Learned (The AI Mindset Shift)

This project changed how I view AI. It’s not just about "having AI do the work"—it's about Engineering the Efficiency.

1. AI for Writing, Not Just Thinking: Don't pay the AI to "think" through every repetitive step. Use it to write a deterministic script that executes the steps. It’s the difference between hiring a driver and building an autopilot.

2. Token Economy is a Skill: Being "efficiently lazy" means finding the path of least resistance. Keyboard shortcuts (Tab/Enter) > Element selectors > Visual recognition. This saved 90% on costs and kept me under my limits.

3. Deterministic vs. Probabilistic: If a process is mechanical, use mechanical tools. Relying on "AI vision" for a 1,000-row loop is risky; relying on a hard-coded tabbing sequence is reliable engineering.

4. Safeguards > Perfection: You don't need a script that can fix its own mistakes. You just need a "Dead Man's Switch" that knows when to stop. This prevents an AI from "going rogue" on your live ERP data.

5. The North Star Mindset: Keep the end result fixed, but stay flexible on the "how." The "3 browser" trick was the spark, but the "Tab" epiphany was the actual fuel that finished the job.

6. Respect the "Cooling Off" Period: Burning through your tokens isn't a failure—it's a signal. That 1.5-hour lockout was actually the best thing that happened, because it forced me to find a more efficient way to work.

More month-end automation updates coming soon!

Special thanks to everyone who commented on Part 1—your insights on "constraint-led creativity" were exactly what I was feeling that night at 00:30 AM!

31 comments