
Series
Build Log · One Person × AI, In Practice
This series has only one rule: write only what I actually did. In November 2026 I went back through these posts and withdrew the figures I could no longer verify, such as per-session costs, and corrected the participant count to what it was: a pre-event expectation. What remains is the method, where it went wrong, and how to check it next time. I am not much of a programmer. What I do is define the problem, break the tasks apart, guard the quality, and wire the tools together. The tools will change; the judgment part, you have to walk through yourself.
01Planning Live Meeting Translation: Separate Checks, Costs, and Data FlowsChecking the source text, translation, costs, and data flow separately; the old cost figures have been withdrawn.
02When AI Starts Choosing Its UsersA Claude troubleshooting log from Shanghai Hongqiao Airport: what a 302 can and cannot prove, and what to ask before relying on a backup.
03Generate Images Without Switching Windows: A Claude Code and Codex CLI WorkflowNo window switching, images saved back into the project; no API key does not mean free or unlimited.
04I Asked Codex to Find Claude's Mistakes—and Brought the Second Opinion Back to the EvidenceWhether a second AI helps depends on keeping it independent, then bringing its opinion back to the evidence.
05A Knowledge Management Pipeline: Let the System Handle Repetitive Organization, Keep Understanding for YourselfMaking collection, syncing, and sorting connect, and making failures visible.
06Can Pilgrimage Footsteps Be Counted Directly as Carbon Reductions?The Baishatun pilgrimage × GPS check-ins: footsteps can be recorded, but not counted directly as carbon reductions.
Newsletter
Subscribe to new posts
Get an email whenever a new article is published. No ads, no list-selling, one-click unsubscribe anytime.
You will get a confirmation email; click the link to complete your subscription.