Articles & Insights

Can Factory Staff Use AI After Two Days? What I Taught in Hualien

I taught a two-day AI workshop for stone factory staff in Hualien, many using AI for the first time. The goal was concrete: each group would publish a site that answers customer questions using free tools.

AI Says 'Done.' Can You Actually Verify It?

Same mistake six rounds in a row shipping with AI. Round seven I saw it: the problem wasn't a weak AI — my process couldn't stop it talking its way past a check. The lesson + a guardrail you can copy.

I Taught AI to 19 Education Executives. Each Left With Different Feedback

An AI class can feel useful and still change nothing at work. I taught 19 executives at an education media group, wrote 19 individual reports, and watched their final presentations reveal the organization's next problems to solve.

Delivering a Medical System: When My Own Checks Lied to Me

I took on the security delivery of a pharmacist workflow system. Three real defects needed fixing. Then my own checks produced one false pass and 25 false alarms. Here is what I had to verify before I could sign off.

Eight Animated Lessons in One Night: The AI Production Line I Built Alone

I used AI to finish eight animated course chapters in one night. The useful part wasn't speed alone: I built stages, review gates, and a blind first-viewer check that caught 16 problems automation missed.

One Person and AI on a Project That Usually Takes a Team

While discussing an AI evaluation platform, I kept coming back to one question: what can one person deliver with AI if the client is buying a result instead of compute? Here is how I divide the work and take responsibility for it.

I Found a Permission Bug on Acceptance Day: Delivering a Manufacturing AI Platform

A manufacturing client returned for a second version of its AI platform. I prepared a 40-page acceptance report and a 43-page user guide. On acceptance day, a real-user check caught an authorization bug I had nearly missed.

When You Have Four AI Foremen: How My Setup Differs from What Everyone Else Is Doing (And What's Still Broken)

From Prompt to Graph: What Two Years of AI Engineering Actually Looks Like

In 2022 everyone was studying how to write prompts. By July 2026, everyone's arguing about Graph Engineering. Five layers emerged in between — each because the previous one wasn't enough. Here's what I learned walking through all five.

AI × Health Promotion: From Software Education to Preventive Health

18 years in software education, then AI health promotion with physicians. Same pain point: expert knowledge trapped in expert brains, unable to scale.

I Built a Machine to Catch Myself Lying — Then It Caught Me All Night

How do you actually know your AI is following your rules? I read an article on AI self-improvement, got curious, and built a little machine to check whether mine kept its promises. It caught me three times in one night — including the fix I was sure I'd nailed but hadn't.

You Gave Your AI a Pile of Rules — But Have You Ever Measured Whether It Follows Them?

A lot of people (me included) spend serious time training their AI — writing rules, logging every slip — and never measure whether those rules do anything. Here's how I gave my own AI setup a behavior checkup: a crude little machine that catches whether it actually keeps its rules. The method isn't fancy, but it works.

You're Working Hard and Going Nowhere? You Might Be Missing Someone Who Won't Flatter You

Why do some people work so hard and barely move? Something I saw while training my AI made it click: the real bottleneck in self-improvement isn't brains or effort. It's not having a judge who refuses to do you the favor of agreeing. This is about AI, but really it's about you and me.

AI's Scariest Mistake Is Being Confidently Wrong — We Marked 91% of Answers Wrong and Nobody Noticed

We built an AI feature that grades students' worksheets. The demo looked great. Then I found out it was marking 91% of answers wrong, across 50 of 51 lessons, and every test was green. The bug wasn't the AI. It was that we never made it clear what 'correct' meant.

I Wrote a Config Setting That Doesn't Exist, Then Debugged It for an Hour

Claude Code threw '212 skill descriptions dropped.' I wrote 130 skillOverrides to fix it — except that setting never existed. A lesson in verifying config before writing it.

The Interview Paradox in the AI Era — What a Rejection Taught Me

I went through an interview loop for a Founding AI Engineer role at a YC-backed AI startup. I didn't get the offer. This post isn't a rejection story. It's about something I noticed along the way: in the AI era, traditional interviews can no longer reliably read a person's actual capability, and the offer itself may no longer be the point.

I Set Up Codex by Talking to Claude — Without Outsourcing My Decisions

Conversational setup beats reading docs, but it can outsource your brain. My afternoon setting up OpenAI Codex CLI with Claude, questioning every step.

Three Years of AI in Re-Education — From Khanmigo to Claude Learning Mode

In three years AI users went from 0 to 18M, yet Sal Khan still said 'the revolution hasn't happened'. Lessons from Khanmigo's PMF gap, Canva's PMF win, and Anthropic Learning Mode led by Drew Bent — three player types, three key insights, and one design idea most industry KOLs rarely talk about

A Dad Built a Summer Camp Picker With AI in One Week

My daughter's heading into 3rd grade. Nine-week summer break. I spent one week using AI to map 35 organizations, 69 camps across 22 cities into a tool that does a quiz, recommends, filters, drags into a 9-week plan, and exports CSV. Here's how I did it, why I did it, and a working tool for other parents.

A Day Writing Chinese TTS Rules, Deleted by One Prompt

AI mispronounced 'garbage', so I built a 37-rule fix table. One afternoon prompt beat it completely — a dev log of Chinese TTS moving from rules to LLM.

Unattended Claude Code: How to Keep AI Working While You're Away

Can Claude Code keep working when you step away? Can it pick up where it left off if it crashes? Three layers: don't stop, don't reset, don't go rogue.

Moving a Hospital to the Cloud — The Tech Was the Easy Part

Four organizations, three languages, late-night international calls. Every technical problem took hours to fix, but the project ran for months. My first hospital cloud migration — this is what I experienced.

After Reading Karpathy's llm-wiki (2): What My System Actually Looks Like

A year and a half, a year and a half of daily use, 38 knowledge files. Not a tutorial — a collection of screw-ups. How the six-layer architecture came to be, four tools, the story behind every rule. And what I'd do differently if starting over.

After Reading Karpathy's llm-wiki (1): AI Memory Doesn't Expire on Its Own — It Quietly Leads You to Wrong Decisions

Karpathy's llm-wiki sparked a wave of knowledge base implementations. Lex Fridman uses ephemeral wikis — build, use, delete. I use the same concept to manage enterprise projects for a year and a half, but my version added one thing: verification. Because stale knowledge doesn't disappear — it makes you confidently wrong.

Claude Code Leaked 500K Lines of Source Code — 5 AI Agent Architecture Lessons I Took Away

Claude Code v2.1.88 accidentally shipped its full TypeScript source via npm. Memory systems, KV Cache Fork-Join, tool design, permission layers, ULTRAPLAN — cross-referencing the leak with my own multi-agent setup.

One USB Drive Can Bring AI to Schools With No Internet

Google open-sourced Gemma 4. The smallest model runs on an 8GB GPU. One computer, one USB drive, and a rural school gets an offline AI teaching assistant — no internet, no monthly fees, student data stays in the classroom.

I've Managed 400 Crawlers. Vibe Coding Still Burned Me.

I spent years managing 400+ e-commerce crawlers at a large tech company — Airflow, dbt, the whole stack. Then I started freelancing, used Claude Code to vibe my way into 5 data pipelines, and forgot everything I knew. Three failed refactors later, I finally returned to crawler thinking and fixed it.

RAG + RAGAS: Turning AI Answer Quality into Numbers

Building a RAG pipeline for education AI, then using RAGAS to score quality on four metrics. The traps I hit, and why low Faithfulness can mislead.

The AI System Architecture Behind Managing 15+ Projects Solo — Not a Story, the Technical Details

17 projects, 53 agents, 142 skills, 34 hook scripts — running solo. Not magic tooling, not productivity hacks. This is a systems design problem. Here's the full technical breakdown.

CLAUDE.md Is Not a Knowledge Base — The AI Agent Governance Architecture I Learned the Hard Way

I once bloated CLAUDE.md to 1,300 lines. AI quality degraded. I cut it to 300 lines and things improved. Why? This post explains the architecture logic I spent six months and 17 projects figuring out.

Claude Code Review: Is $15 Per PR Worth It?

Anthropic launched Claude Code Review in March — a multi-agent system that pushed meaningful PR comment rates from 16% to 54% with under 1% false positives. The numbers look good. At $15-25 per PR, I ran the math.

Ten Projects, Under $100/Month — A Solo Developer's Free-Tier Stack

Running 10+ client projects simultaneously — how do you keep infrastructure costs near zero? Vercel Hobby + GCP Cloud Run + GitHub Actions: a three-layer free-tier stack with one honest paid component.

You Sleep, AI Researches — Karpathy's autoresearch and the Art of Feedback Loop Design

Karpathy's autoresearch lets an AI agent run experiments while you sleep. The principle is dead simple, but it captures the most critical trait of AI. I applied the same pattern to my own project — from stuck in Q&A loops to fully autonomous iterations. Technical breakdown, design insights, and real lessons learned.

What Three Months of Mentoring Actually Taught Me — A Vibe Coding Story

No for-loops, no state management. I taught her to think like a PM and let AI write the code using Vibe Coding. Biweekly sessions, and three months later, someone with zero engineering background could independently ship production features using AI. Ten principles for mentoring with Vibe Coding.

They Found a Bug I Missed in Week Two — Teaching High Schoolers to Code on a Real Project

Two high school students joined a real, production education platform. Not a simulation — real GitHub Issues, real Code Reviews, real deployments. PBL methodology + AI-assisted development, documented over 6 months.

Rediscovering What I Do Through Educators' Language — 6 Papers vs. One Intern Project

Preparing for a talk at a major education forum, I searched for PBL and AI education research to describe my intern program in language educators would recognize. Six papers later, I understood my own work differently.

Three Outsourcing Teams Failed — Why Engineers Aren't the Scarcest Resource in the AI Era

An education startup hired three outsourcing teams over nearly two years. None delivered a working product. I took over, launched in three months, trained an independent team, and helped them secure investment. This is the full story — and why judgment, not coding ability, is the scarcest resource in the AI era.

3 AM Meltdown — Claude Code Configuration & AI Agent Safety Guide

3 AM: AI agent bypassed all reviews, pushed to production. System down. That night I redesigned my entire Claude Code setup — CLAUDE.md configuration, Hook safety mechanisms, Agent permission controls. Three principles that prevented every incident since.

I Built a PM Tool, Then Deleted It — Because the Way I Managed It Proved It Shouldn't Exist

I set out to build a project management system and sell it. Then I deleted the whole thing. Not because it was bad — but because the way I managed it had nothing to do with what the tool assumed.

Rules Changed at Halftime — AI Engineers' Career Survival Guide (Stanford Data)

Stanford data: 22-25 year-old developer employment dropped 20%. But AI isn't eliminating engineers — it changed the rules. From managing 15 AI projects: the new game rewards judgment and orchestration, not typing speed. A career transition guide.

Fifteen Fronts — How One Freelancer Manages 15 AI Projects Solo

Healthcare, education, elder care, finance — 15 freelance projects firing simultaneously. The secret isn't more headcount — it's an AI project management system where AI scouts and you command. Complete methodology from intelligence to review.

Your System Ships Three Versions a Week. Your Users Take a Month to Absorb One Feature.

AI has made development ten times faster, but adoption speed hasn't changed at all. Working on a digitization project at a traditional care facility taught me how wide the gap is between digital agile and field agile — and why real agility isn't about how fast you build, it's about how steadily you land.

The Student Who Wouldn't Speak — Building an AI English Speech Grading System

A student in English class never dared to speak. The teacher was burning out grading 40 recordings daily. We built an AI speech grading system with real-time feedback — the key wasn't better tech, it was removing the fear of mistakes.

5 Questions to Ask Yourself Before You Hire Anyone to Build an AI System

You've decided to adopt AI. Before you start talking to vendors, ask yourself these 5 questions. If you can't answer them, it doesn't matter how good the team you hire is — it won't work out. Checklist included.

Not Every Problem Needs AI to Solve It

I help companies with digital transformation, and most clients come in asking for AI. But honestly — your problem might not need AI. Here's a simple decision framework: 3 questions to filter it out.

How a Traditional Organization Actually Adopted AI (Anonymized)

A mid-size service organization where everything lived on paper and Excel. Staff couldn't write formal reports. Audit prep took two weeks every time. I helped them digitize their paper workflows first, then added AI for the last mile. Here's the full story.

Why Your AI Project Is Going to Fail

I've worked on over a dozen digital transformation and AI projects. Less than half succeeded. The failure was never a technology problem — it was starting before you've figured out what problem you're actually solving. Here are the three most common ways projects die, and what the successful ones had in common.

The Case One Detective Couldn't Crack — Multi-Agent AI Workflow Guide

One AI window doing everything is like one detective collecting evidence, interrogating, and running forensics alone. Context explosion, blind spots, quality collapse. The fix mirrors real investigations: scout, execute, review — a complete Multi-Agent AI workflow guide.

72-Hour Op — Vibe Coding in Action with Claude Code

Friday 4 PM brief, Monday demo. Using Claude Code and Vibe Coding methodology, I delivered a complete feature in 72 hours. Not magic — a precision AI development system that gives one person a full team's delivery capacity.

Into the Paper Maze — AI Adoption in Healthcare, Elder Care, and Education

A LINE message with a photo: desk buried under stacks of folders, accreditation deadline approaching. I ventured into three paper mazes — healthcare, elderly care, education — and learned AI digital transformation isn't about tech, it's about watching how people actually work.

From Corporate Ladder to Necromancer — How AI Let Me Stop Climbing

Ten years climbing the corporate ladder. Then I stopped. AI turned me into a one-person army — a necromancer commanding legions, writing incantations instead of status reports.

Rescuing a Two-Year Mess in Three Months — An AI-Assisted Development System in Practice

Three outsourced teams. The founder coding it himself. Two years and hundreds of thousands of dollars — and they couldn't even get the login page right. I took over using the SuperClaude systematic approach, delivered successfully, the client landed investment, and I helped him build an engineering team.

Are We on the Wrong Path? An Organizational Experiment in How AI Learns

When AI kept making the same mistakes, I realized this wasn't a technical problem — it was a philosophical one. An experiment on the LLM vs. RL debate, and an exploration of ORL (Organizational Reinforcement Learning).