Where I Stand
- Karpathy went from 80/20 to 0/100 on writing code vs. delegating to agents. I'm using AI heavily — Claude Code, skills, agents — but for my own workflow, not for my customers. The AI investment is flowing inward, not outward.
- SupportWire is positioned as "AI customer support." But is the AI actually built, or is "AI" a positioning word on a landing page for a product that's still pre-launch? Karpathy shipped Dobby the elf claw in three prompts. What's SupportWire's equivalent?
- Karpathy's AutoResearch beat two decades of hand-tuned hyperparameters overnight. The lesson: stop being the bottleneck on things that can be automated. What am I still hand-tuning at FeatureOS that an agent could optimize — pricing experiments, onboarding flows, churn prediction?
- I'm building an elaborate vault and AI skill system to organize my thinking. Karpathy would call this infrastructure for the CEO, not infrastructure for the product. The tokens I'm spending on vault-building could be spent on product-building.
- "Everything is skill issue" — Karpathy's framing means the bottleneck is always my ability to orchestrate, not the tool's capability. The tools are ready. Am I orchestrating them toward revenue, or toward personal productivity theater?
- The agentic web means apps become APIs. Is FeatureOS ready for a world where AI agents submit and triage feedback on behalf of PMs? If not, a competitor who thinks this way will eat the market.
What I Need to Change
- Ship one AI-powered feature in FeatureOS within 30 days. Not "explore AI." Ship it. Auto-categorization, smart deduplication, sentiment analysis — pick one and put it in customers' hands.
- Define SupportWire's AI capability in concrete terms: what does the AI do on day one that a human can't? If I can't answer that in one sentence, the product isn't ready to launch.
- Redirect 50% of the time I spend building personal AI tools toward building AI features for paying customers.
- Audit FeatureOS's API surface. Is it agent-ready? Can an AI agent create, categorize, and prioritize feedback via API? If not, build that before a competitor does.
The Hard Question
You're spending more tokens organizing your own thinking than building AI into the products people pay you for — when did personal productivity become more interesting than customer value?