Since 2023 I’ve spent over $1,000 on paid AI tools—from API credits to the first months of ChatGPT Plus and a couple of first-comer apps. The goal was to use my spare money as a learning budget: learn how AI works, how it thinks, and what to do with it.
The grind I wanted to escape
B2B sales back then meant personalizing icebreakers manually. I’d open each site, qualify the company, find the right lead, then write the right copy to personalize it. Upload it to an outreach tool and send a batch. It worked, but it took hours.
Internet Diet is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.
I knew automation was cool and I could work my way around Google Sheets and basic HubSpot workflows—but I had the same bottleneck every time I came to work. Manual data enrichment was all over the place and data providers were expensive.
Phase 1: spreadsheets as training wheels
My first leverage wasn’t “AI.” It was Sheets.
Joins:
VLOOKUPto match companies, domains, and emails across tabs.Parsing:
REGEXEXTRACT()to grab IDs, domains, and fragments from messy text.Light pulls:
IMPORTXML()to collect structured bits when it was stable enough.
Those were my small hacks back then. The real unlock came from opening Inspect in the browser, because during “scraping,” half the time the data I needed was quietly exposed as JSON on the Network tab. I stopped dumb scraping and called the endpoint from a script instead.
About those prompt marketplaces
I browsed them. I didn’t buy. Or I waited for someone to leak them :)
Templates gave me structure, but results came from context. The framework I still use is stupid-simple: Context → Task → Output shape → Constraints → Few-shots.
3.5 vs 4 (my 2023 reality)
ChatGPT 3.5 was great for editing and Excel formulas. It fell apart whenever I had larger files.
Then GPT-4 landed (March 2023), and the longer context and more IQ meant I could attempt things that used to be “nah, it’s too dumb for that.” I wasn’t suddenly a software developer, but I could get some code rolling.
During that year, browsing in ChatGPT was a moving target—disabled for a bit, then re-enabled in late Sept 2023—so I leaned on Perplexity when I needed quick links and kept my coding loop separate. That worked fine.
Phase 2: Apps Script, buttons, and one-page UIs
This is where “vibe coding” started to feel possible.
I could take an idea and ask the model to draft Google Apps Script that would:
fetch the data via
UrlFetchApp,normalize it, and
write rows into a target Sheet.
Now I could run a formula like an app on Google Sheets and scrape a site automatically. The time from idea to barely working code was usually 3–4 hours of back-and-forth prompting: copy-paste, run, error, fix. It was tedious, but it worked.
The true value at work
At carVertical I had to do sales outreach to car dealerships at scale: car ads, websites, contact info. Think 10k–13k pages. I opened DevTools, found clean data sources, and asked GPT-4 to cook an Apps Script that pulled, parsed, and wrote everything into Sheets.
Later, my role drifted from sales rep into sales ops—HubSpot, cold outreach tech stack (email infra). GPT-4 didn’t replace judgment; it raised the ceiling on what I could try in one workday.
Tools that actually helped (in that 2023 window)
Perplexity because ChatGPT didn’t have web search back then.
TypingMind because I could bring my own API keys, set profiles/system prompts, and get a cleaner workflow than the vanilla chat UI at the time.
Instant Data Scraper—a Chrome extension that taught me how web scraping works.
What changed (and how I measured it)
I stopped measuring progress by vibes or likes and started measuring project complexity:
Editing → Formulas → Apps Script → Web Apps
The curve bent upward once I was willing to pay for the frontier models and use them as much as I could.
The take I’ll defend
There’s no magic prompt. You can buy a course. A playbook. It won’t solve your personal problem. You read, try, and see what works. Then take it and move on to the next problem.
What’s next (and what I’m not covering here)
This is Part 1. I’m deliberately stopping before the era of polished search/browse in the chat UI and the agentic IDE boom. That comes in Part 2.
Internet Diet is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.