This is the English companion note for a longer Chinese essay. The full version is available on the Chinese site.
In early April, I did not think of myself as a developer. I started because I saw that AI could help ordinary people turn ideas into working systems, and because my own study process was missing something important: a real review loop.
What began as a small learning tool became a broader question: how can a person preserve thinking, review mistakes, define problems, and keep learning over time?
The answer did not come only from technology. It came from years of reading, from Dabie Mountain reading programs, from education work, from a visit to Hu Shi's hometown, from Steve Jobs's idea of connecting the dots, from Charlie Munger's discipline of avoiding stupidity, and from Duan Yongping's insistence on doing the right thing and correcting mistakes quickly.
Over roughly 35 days, I built six systems:
- Socrates Focus, an evidence-constrained thinking and learning coach.
- IELTS Assistant, a live learning tool for reading, writing, OCR, and review cards.
- Reading Reflection OS, a system for turning reading into long-term memory.
- ShadowBuyer / Booking Audit, a public-facing evidence audit experiment around platform data and claim boundaries.
- Solo AI Company OS, a Markdown operating system for coordinating AI workers.
- SACP / AgentOps Doctor, a protocol and reference tool for auditable AI agent work receipts.
The projects look separate, but they share one underlying question: in the AI era, how can one person save thought, organize action, verify results, and keep improving after failure?
The most important shift was not learning to ask AI for code. It was learning that AI reduces implementation friction, but it does not remove the need for judgment. A person still has to define the problem, break it down, verify evidence, decide boundaries, and correct the system when it drifts.
That is why auditability became central. After long conversations with AI, I saw context drift, hallucinations, and forgotten goals. A useful agent should not merely say "done." It should leave evidence, scope, decisions, responsibility, and a reviewable receipt.
My current belief is simple: the future belongs not only to stronger models, but to vertical agents, durable memory, evidence-bound workflows, lower-cost local or specialized models, and humans who can still define problems clearly.
From beneficiary to builder, this is the thread I am trying to follow: technology should help people become more complete, not more dependent.