Ideas I Shelved in 2013
OpenAI ran a Super Bowl ad this year: you can just build things. It took a couple months but the copy from this ad finally sunk in and reminded me of this great piece, The Cult of Done, written by Bre Pettis and Kio Stark from 2009.
Bre and Kio’s wrote a 13 rule manifesto, but three of them stuck with me the most: start on an idea immediately, treat your projects as perpetual drafts, and call it done even if it fails. Because you learned something along the way.
At the time, the Cult of Done inspired me and frustrated me. As a non-engineer in operations, a part of me felt stuck in tools and jobs that held me back from what I could build. I had some inclination to learn programming, and I’m proud of what I built in my career anyway. But looking back now, I wonder what I could have built with AI in hand.
The Cult of Done was always right. For non-engineers in ops it was also aspirational. Shipping something rough up the chain was risky. Rough-looking things got sent back even when they worked and the solid working version required skills/time we didn’t have.
AI has changed this. Anyone with a good idea (with missing skills) can describe what they want in English and a solid working version appears. And going from idea to impact is insanely fast.
In 2013, I had the ideas but I didn’t have the skill to make them a reality. I remember thinking wouldn’t it be useful if we could automate the supply plan process? and then getting pulled back into whatever fire was burning that day. Or what if we could build something that allocated shipments to the highest-priority customers on its own?. Nice ideas and all, but they needed complex ERP modules and/or dozens of man-hours per cycle to solve.
Last year, AI was mostly an advisor to us at VergeSense on supply-chain questions, especially the geopolitical ones. How do I import into Egypt? How can we minimize GST for our customers in India? Which HTS codes are hit by the new tariffs? The things that used to require a trade consultant or a week of research came back in seconds from an LLM.
But this year, thanks to advances in Claude Code, OpenClaw, and Codex, I’ve started building. The first thing to really have an impact was the agent itself. We quickly set up something that turns meeting notes into a Confluence page. Essentially a scribe that automatically writes key decisions down. This meant no more going back a week later asking wait, what did we agree to? What used to take an hour of writing takes two minutes of review. The same agent searches the rest of our knowledge base from Slack. This stack is becoming a consolidated & default way to search our knowledge base (vs. scattered searches around Google Drive, Slack, email, etc.). We built another tool that tells us when customers stop using the expensive devices we’re still paying a subscription on. That analysis used to live in a spreadsheet and cost a few hours every cycle to scaffold up. Now I forward the agent an email and it does the rest.
None of these are perfect. They’re rough but they work. The win here isn’t these exact tools, and I’m not trying to prescribe a boilerplate solution, but the win to me is that I have stopped treating “could we automate this?” as a question for someone else. Instead, I just start building.