Why I built Coop.ai

The moment that triggered it was small and stupid. I was sitting in front of a job posting at 11pm with eight tabs open — LinkedIn, the company site, Glassdoor, Blind, two news articles, my resume in Google Docs, and a ChatGPT window where I'd pasted my background for the fourth time that week. I was about to ask the same question I'd already asked about three other companies that day: "is this actually a fit, or am I forcing it." And I realized I'd been doing this for weeks. The same setup. The same paste. The same context rebuild. Every single time.

The spreadsheet was the second insult. I had a Google Sheet with columns for stage, next steps, comp, key contacts, drop-down statuses — the works. It was beautiful for about a week, then it started rotting. By week three I had no idea what stage half the companies were actually in. I'd had a meeting with someone at one of them and the row still said "applied." The sheet was a snapshot of intentions, not a system that kept up with reality.

So I started hacking. The first version was just a Chrome extension that detected the company on the page and pulled in firmographics on whatever was in front of me. Tabs collapsed into one click. Within a few days I added save-to-list. Within a week I had a Kanban. Within two weeks I'd wired up Gmail so emails would attach themselves to companies by domain, and Granola so meeting transcripts would do the same. I was building it the way you build something for yourself: no roadmap, no users, just the next thing that annoyed me.

The real unlock came later, and it wasn't a feature. It was the realization that the AI part — the chat I'd bolted on as an afterthought — was the whole point. When I gave Coop, the advisor inside the extension, full context — my profile, the company data, the job description, the emails I'd already exchanged with people there, the meeting transcripts from calls I'd already had — the conversations changed completely. I stopped explaining and started asking. "Given what we already talked about with these folks, what should I lead with in the application?" And Coop would actually know. Not from a prompt I'd just pasted, but from data the extension had been quietly assembling the whole time.

That's the part the spreadsheet and the bookmark extension and the standalone ChatGPT window can never give you. It's not the research, it's not the Kanban, it's not even the scoring. It's that the advisor has the same picture you do. The full timeline. The actual relationship history. You stop being the integration layer between four tools and a model.

The thing it taught me about building with LLMs is the bit I keep coming back to. I had Coop's opinions hardcoded for a long time. Salary floors got rewritten in the prompt as "WALK AWAY BELOW $X." Work arrangement mismatches got an automatic three-point penalty in the post-processor. Dealbreakers got severity-scaled in code. The result was that when I asked Coop to draft a cover letter for a role slightly below my floor, he'd give me a six-paragraph fit lecture instead of a draft. The model was being told, by my code, to refuse before it ever saw my actual question.

I ripped all of it out. Every interpretation Coop applies to my data now lives in a single editable textarea in settings called Operating Principles. The code emits neutral facts — "base salary floor: $100K" — and the textarea tells Coop how to read them. If I want him sharper, I rewrite the textarea. If I want him to never refuse a draft request, I write that. No code change, no deploy.

Coop is in private beta. I built it for my own search and use it daily; if you want in, the beta access page is a real form. The reason I wrote any of this down, though, is that the principle above is bigger than the tool itself. Coop is the first place I tested it end-to-end. It's the one I'll carry into everything I build next.