All Articles
Articles grounded in projects, measurements, and public records. Browse by topic, find the answer first, then follow the original sources to verify it.
No. Playwright is a browser automation and testing tool. WebMCP lets a cooperating website declare capabilities hidden behind its interface as structured tools. They solve problems at different layers.
Read articleEvidence and constraints behind choices in Taiwan transport, popular products, major events, and seasonal travel.
A 2026 transportation study examined Tainan roads where mandatory two-stage left turns for scooters were removed. Crashes did not rise: the estimated number fell 21%, while the number of people injured fell 19%. Those numbers make a strong headline and an easy overclaim.
Read articleThe Owala FreeSip did not go viral because vacuum insulation suddenly changed. Its appeal is a lid that is easy to demonstrate: sip through the straw or tilt the same opening for a faster drink, with a cover and carry loop built in.
Read articleThe 2026 Aichi-Nagoya Asian Games run from September 19 to October 4. Japan is one hour ahead of Taiwan, so convert official schedule times. Events, entries, and broadcast channels may still change before competition.
Read articleAs of August 13, 2026, the Japan Weather Association has not released its 2026 autumn-color forecast. Tenki.jp says the new season’s information is expected around September. Precise “2026 peak dates” circulating now are generally not an official seasonal forecast.
Read articleMethods, frameworks, and hard-won lessons from putting AI into a real development workflow — from knowledge management to tooling.
No. Playwright is a browser automation and testing tool. WebMCP lets a cooperating website declare capabilities hidden behind its interface as structured tools. They solve problems at different layers.
Read articleAI writes code ten times faster than humans, and it creates hidden technical debt ten times faster as well. It excels at delivering code that compiles and passes tests, but secretly slips in any, ignores edge cases, and reinvents existing helpers.
Read articleLately there is a pile of "markdown for AI": llms.txt, AGENTS.md, CLAUDE.md, OKF. They keep getting lumped together, so people ask "if I have one, do I still need the others?"
Read articleI built this site piece by piece with Claude Code and Codex. Every session starts cold, so they do not know what came before. I used Google’s Open Knowledge Format (OKF) to turn the repo’s "brain" into a structured Markdown knowledge base.
Read articleThe phrase "OKF replaces the vector database" is confusing because it pits three things from different layers against each other. OKF is a format for knowledge, RAG is a workflow for using it, and a vector database is one retrieval infrastructure.
Read articleRAG often looks great in a demo, but it starts to fail when the knowledge is structured, frequently updated, and full of relationships. The trouble is usually not in the model; it is in the retrieval layer.
Read articleThe OKF spec is deliberately tiny: a directory of markdown, YAML frontmatter, three rules — that is it. This piece walks through a minimal, runnable bundle.
Read articleOnce the three layers are clear, the practical question is how OKF and RAG compose. The useful architecture has two layers: OKF as the source layer, vector/RAG as the retrieval layer.
Read articleThe most damaging moment in a subscription product: the user pays, returns to the app, and the screen still says Free. The money moved; the entitlement did not. This is not a purchase-flow bug — it is a state synchronization problem.
Read articleAn AI agent script that passed a hundred manual runs died on its first night under launchd. And not the noisy kind of death — the kind that prints an error message and then returns exit 0.
Read articleNotebookLM has no public API, so I built one with browser automation and published it to PyPI. Getting it to work took a weekend — the real work was the next two releases of reliability fixes.
Read articleA five-player game of Avalon traditionally needs a matchmaking service, a state store, a WebSocket server, and a cleanup job. Built on Cloudflare Durable Objects, those four things collapsed into one call to idFromName and one setAlarm.
Read articleA summary makes material shorter without deciding what should happen next. A convincing idea may deserve only a saved source, further investigation, a public article, or a product experiment.
Read articleAfter Google Cloud introduced the Open Knowledge Format (OKF), claims like “Beyond RAG” and “the vector database is being replaced” spread quickly. The three tools solve problems at different layers.
Read articleAI development began as an exercise in asking better questions: refine the prompt, add examples, and specify the output. Once models started reading repositories, using tools, and working for longer, the true unit of engineering was no longer a single instruction.
Read articleWhen I first used AI to build apps, my most common instruction was: "Finish this feature for me." Sometimes it worked so quickly that I thought I had found a new way to build software.
Read articleNotes should be captured when an idea appears. The decisive issue is not how smart the AI is, but whether thinking and recording are forced into separate places.
Read articleGoogle NotebookLM is good at turning documents into an interactive knowledge base, but it has a practical limitation: it cannot be controlled directly from a program.
Read articleAI comparisons often focus on features and price. I wanted to ask a different question: how differently would three mainstream models answer the same technical prompt?
Read articleFirst-hand testing of models and tools: which one fits where in your workflow.
To turn a YouTube video into structured notes, the intuitive pipeline is: scrape the transcript, feed it to a model. You do not need any of that — Gemini's fileData accepts a YouTube URL directly, and the model watches the video itself, subtitles or not.
Read articleI had barely finished comparing Gemini 3 Pro, GPT-5.1, and Claude Opus 4.5 when the version numbers jumped again. Google released Gemini 3.1 Pro, OpenAI moved to GPT-5.6, Anthropic introduced Claude Fable 5, and xAI brought Grok 4.5 into the fight.
Read articleThis was an experiment in designing an agentic workflow around Google tools. I deliberately limited the available services to find the practical ceiling of the Google toolchain.
Read articleAI is no longer a distant idea from science fiction. We do not have to wait for AGI or household robots to see changes in everyday behavior.
Read articleGemini 3 Pro, GPT-5.1, and Claude 4.5 Opus arrived close enough together to invite comparison. The differences were no longer limited to model size or response speed; they also appeared in reasoning, context handling, and the work each product was built to support.
Read articleGetting started with Three.js, WebGL, and Blender, alongside playable experiments in the lab.
Godot 4 games easily hit 60fps on Mac or PC, but clicking Export HTML5 and dropping the build onto a web server often leads to an instant blank screen: gray views, cross-origin security barriers, iOS crashes, and audio deadlocks.
Read articleFloor plans are the most universal language of space, and most people cannot actually read them — between that 2D drawing and "what it feels like to live there" sits a translation gap. I turned a real three-bedroom layout into a browser 3D space with orbiting views and draggable furniture.
Read articleMy Sokoban game has two solvers. The algorithm is identical — plain breadth-first search — one written in GDScript inside Godot, one in TypeScript running in the browser. This is not wheel reinvention, because they are two entirely different products.
Read articleI curled my own article page, stripped the scripts, and counted what was left: 46 characters. The title. Not one character of body text. All 34 article pages looked like this — to anything that does not execute JavaScript, the site was nearly blank.
Read articleBoxCat launched with 40 levels. For version 1.1.0 I wanted 100, with no unsolvable maps, no invented par scores, and no sudden drops in the difficulty curve.
Read articleOpening the network panel on Midnight Run: 1,894 KB transferred, of which 1,714 KB is two GLB models. On the same page three.js is 564 KB on disk and only 142 KB over the wire.
Read articleMidnight Run does not use a track I drew. It runs on a real block of Taipei's Xinyi District. The route comes from an OpenStreetMap snapshot, the city is generated in Blender, and what reaches the browser is a 1.3 MB GLB.
Read articleThe day I finished my typing test I ran it once myself. Speed, accuracy, passage changes — all fine. The next morning I switched back to Zhuyin and the screen turned red. I pressed ㄐ and the app immediately decided ㄐ was not 今. I had missed it the day before because I had tested in English.
Read articleWebGL is a low-level graphics API built into the browser. Through a canvas, JavaScript can send vertices, colors, textures, and shaders to the GPU to draw real-time 2D and 3D without Flash or another plug-in.
Read articleThree.js lets a website become more than a flat page. It wraps low-level WebGL in approachable ideas such as scenes, cameras, materials, and lights, making it practical to build 3D product views, data visualizations, games, and interactive stories with JavaScript.
Read articleBlender is a free and open-source 3D creation suite. Modeling, sculpting, materials, animation, simulation, rendering, compositing, and video editing can all happen in one application. It can also become an automated asset factory for the web.
Read articleFirst-hand notes from idea to launch: what got built, why, and what it taught.
I have a row of "watch later" YouTube talks, each one easily half an hour, each one still unopened. Saving is quick; turning it into something in my head is the cost I keep deferring.
Read articleThe "chart" you see in the App Store is your country's chart. The United States, Japan, and Taiwan each have their own — not translations of one ranking, but three independent lists. Apple has never offered a global chart at all.
Read articleA solo product has no PM to set priorities, but the product-management work does not disappear. Who needs a feature, why it matters now, and when to stop all return to the same person.
Read articleWhen AI returns an empty response, times out, or breaks the expected format, a product knows to stop. The harder case is a fluent, valid response that drops a constraint or makes the wrong judgment.
Read articleWhen an article looks important but there is no time to read it, saving it for later feels reasonable. Nothing is discarded, and we do not have to admit that we may never return.
Read articleKofAI Note 1.0.0 is now on the App Store. Its current promise can be compressed into one line: “Capture in one tap, and let AI read it for you.” That explains what the product does without defining what the user ultimately gains.
Read articleBabyDay did not begin with a product plan. Our second child was on the way, and the records from our first were scattered across LINE, phone notes, and paper. When we tried to look back, none of it lined up.
Read articleI honestly never planned to build a Pomodoro timer.
Read articleStarting on Medium felt unfamiliar. I was used to expressing ideas through code, and an empty writing editor was harder than I expected.
Read articleThe Chinese typing daily challenge — everyone types the same passages today. See where you rank.