> Page 23 My Turn > Page 24 View from the Repair Bench > Page 25 Women of NAMM
AI IN MI I BY PETER DODS DON’T LET AI HELP YOU FAIL
I n my last column, I promised we’d dive into “level four: the workflow automator,” mov- ing beyond chatting with AI to connecting it with systems like Google Apps Script. I have built these systems, and, trust me when I say, they’re game- changers. But the learning curve was brutal. I realize now that if you attempt “level four” without understanding how LLMs “think,” you don’t get automation. You get a Rube Goldberg machine of code. Before we become automators, we must address a trap called contextual inertia. If you don’t understand this, you’ll spend weeks polishing a turd while the AI hands you the polish. A three-minute nightmare I recently experienced was I tried to “auto- magically” sync massive Excel attachments from Gmail to Google Sheets. The files were huge, and the script kept timing out. I asked ChatGPT, Claude and Gemini for help. They dutifully wrote complex “batching” logic, processing 1,000 rows at a time, saving state and restarting after crashes. We even resorted to run- ning it at night while Google’s servers were less busy, but it still crashed sometimes. It was slow, fragile and a nightmare. Frustrated, I finally asked a “stupid” question: “Should I write multiple scripts to process the file column-by-column?” I outlined a manic schedule: I could trigger columns A–C to
three weeks ago? Because LLMs are designed to be compliant, not critical. When you paste a block of code and ask, “How do I fix this error?” AI analyzes that specific block. If your code uses a slow, ama- teur loop, the AI thinks, “I’ll help make this the best loop possible.” It doesn’t think, “This user should ignore their code and use the pro- fessional standard.” The hard truth is that AI will happily help you optimize a bad idea. This applies to anyone using AI. Here’s some examples: 1. The Over-Engineered Itin- erary. You’re planning a European vacation. You ask for train routes between 10 small towns. AI helps you navigate complex transfers for an hour. It never mentions that a $40 rental car saves you 20 hours because you asked for trains. 2. The “Infinite Revision” Email. You ask AI to “soften” a difficult email to your landlord. It tweaks adjectives but leaves the email 500 words long and defensive. It never suggests that a phone call would actually solve the problem. HOW TO PROMPT FOR ARCHITECTURE Reaching “Level 4” means mov- ing past syntax. Stop asking the AI to fix lines of code, and start asking it to build the blueprint. Prompt for the architecture, not just the implementation. First, there’s the “blank slate” prompt: Don’t paste your work- flow — that locks the AI into bad logic. Describe the goal without
the method. Also, don’t ask it: “How do I make this loop faster?” Instead, ask more specifically: “I need to move 50,000 rows of Excel from a Gmail attachment to Google Sheets. If you were a senior en- gineer, what specific architecture would you use to ensure it runs under 60 seconds?” Then, there’s the “wrong an- swer” technique. Cunningham’s Law suggests that the best way to get the right answer on the in- ternet isn’t to ask a question, but to post the wrong answer. If you suspect a more efficient method exists, try proposing something extreme to bait the experts into correcting you. Finally, ask for the professional standard. AI defaults to hobby- ist coding because it’s easier to explain. You must explicitly ask: “Is this the industry standard? Is there a more advanced system a professional would use, even if it’s
“AI is an incredible contractor, but it’s not normally a great architect.” run from 6 to 7 p.m., then trigger columns D–F from 7 to 8 p.m. and finally columns G–I from 9 to 10 p.m. I was ready to build a fragile clockwork monster just to move some text. The AI panicked. “No! That causes synchronization errors. Just use the Advanced Sheets API.” As it turns out, there was a much faster way to handle the data. By switching methods, I could eas- ily process 50,000 rows without hitting any time limits. My script went from a three-minute disaster to a stable, 48-second breeze. THE TRAP OF CONTEXTUAL INERTIA Why didn’t the AI set me straight
harder to set up?” YOU’RE THE ARCHITECT
AI is an incredible contractor, but it’s not normally a great architect. As we move toward automating our businesses with “level four” workflows, don’t let the AI’s compliance fool you into thinking your architecture is sound. Some- times, you have to be will- ing to break your own mold, because the AI never will. MI
Peter Dods is the owner of Honolulu-based Easy Music Center.
22 I MUSIC INC. I MAY 2026
Powered by FlippingBook