Sketched a framework that reads one problem twice — once by ordinary analysis, once through the sixty-four hexagrams — then argues the two readings against each other. One of the four hexagrams is drawn at random, and that is the point.
The caption under a chart said one pair of numbers. The curve above it drew another. The words were typed by a person and the line was drawn by code, and nothing made them agree.
The drawing skill said the default line was a hand-drawn sketch, meant for rough concepts. It had been drawing finished equipment diagrams that way the whole time. I wrote that rule.
Claude Code reads skills from only three places, and none of them is shared across repositories. So one repo became a warehouse. Here is the whole structure, and the two things that kept not working.
Search engines and AI crawlers do not open a page the way a person does. Four times in seven months, the file they were handed was not the page I thought I had built.
The wafer transfer trainer had fifteen chapters and every one of them ended at "did you press the button". Five graded drills became twenty-nine, and one of those five turned out to be unwinnable two percent of the time.
Eleven days, four games swapped out — including the one that drew the most attention. The invite mechanic is what made a game nobody would forward impossible to keep.
Built a travel-compatibility game for youdam.com. Eight questions, two minutes, four travel types. Alone you get half of it — the other half only opens when someone you invite finishes theirs.
An answer engine folded my brand together with a semiconductor company whose name looks like mine. The fix was not writing more pages. It was filling in what this is, and what it is not.
Plugged the MCP in and ran it end to end. Handing over the tools was not enough. It only felt like a fab once there was a door between inside and out — and it cost speed.
Stood up the equipment technology book, then built a gas panel you operate instead of read. The flow animation ran backwards on one line only, and the cause was not direction. It was speed.
The fab is one dark shell with process modules plugged into it. A bright module inherited the shell colour and went white on white; the sweep found twenty-one more in the diagrams.
Wrote an llms.txt for the arcade so answer engines could quote it. The top line got rewritten three times — and each rewrite was a different answer to what the site actually is.
Deleting a class did not delete its seats. The leftovers kept counting, and the numbers the staff read were wrong.
Handing out sixty codes at the start of a class wastes the first ten minutes. One code on the projector, and everyone gets their own seat.
Deleted the entire EduAtlas application and started over. 47 files, 11,419 lines gone in one commit.
The gateway works. A general AI can list the flows, read parameter ranges, open the equipment panel, and run the process itself.
KOH crystal-plane etching is outside what the model covers. Declared the boundary instead of implementing a guess.
Moved Bandi onto a different brain. The hands — the action bus, the eyes, the console wiring — stayed exactly where they were.
An assistant that proposes actions for staff to approve. Then I made it watch what happened after — and stop suggesting things that cannot reach anyone.
Wiped the blog off youdam.com and put a web game arcade there instead. No sign-up. Three minutes per game. Retargeted it three days later.
Sat down and used the etching simulator like a student would. Three things were not working. Then a textbook check found HBr half-built.
A build passing does not mean the numbers are right. Built a physics invariant check and threw forty thousand random inputs at it.
Two weeks after wiring the verified-recipe warehouse in front of the model, one phrasing walked straight past it. "다층 배선해줘" fell through to invention.
Built a SemiFab-specific visual content engine on top of the video factory. A still-first pipeline cut the cost of one iteration by 25x.
Checked every theory tab against the literature rather than against the code. Two modules were teaching the exact opposite of the truth, and the test that should have caught a third was comparing the code to itself.
Every scenario deliberately leaves one step out. The student has to notice what is missing instead of following a recipe to the end.
How do you test a layer that is not deterministic? Not by asking another model. By comparing against what the fab can see.
Replaced the blog landing with an interactive SemiFab DSE demo — HBM physics packs, quantum fabrication tracks, decision gates.
The operations assistant answers when asked, and every answer costs a call. So the always-on view got built without any model at all.
Built the instructor side. What the class is doing, where they are stuck, and what Bandi did about it — all reported into one place.
Not every institution will run the AI tutor. Built a deterministic status board that watches the process and says nothing more.
If a well-known process is asked for, hand over a verified golden recipe before the model gets a chance to invent one.
Finished the basic mode of Bandi, the AI tutor. It reads the live process state — etch depth, sidewall angle, contamination — and helps from there.
Hand-written tests only catch what you thought of. Built fuzzers that sweep the render space from many angles and report a coverage cloud.
Built an operations assistant for the people who run the classes. It borrowed the tutor's shape — same chat, same cursor, different job.
Let students talk Bandi into running the equipment. It lied about having run it. Killed the whole path the next day.
Rebuilt the topic modules as an actual library — shelves, spines, a reading room. Then swept every deck for the errors the redesign exposed.
Westworld became the control and setup center. Declared the three layers of verification, and made measurement honesty a permanent gate.
Split Bandi in two: rules choose the moment to speak, the model chooses the words. Canned lines became the fallback.
Rewrote the SEM cross-section engine as a z-buffer rasterizer. Occlusion became automatic and a pile of culling hacks disappeared.
Wrote down what makes a tutor feel natural: silence, variation, mood inertia, remembering, and admitting limits.
Built a dedicated PECVD equipment scene — parallel plate electrodes, gas plumbing, RF/LF power. Then retired the schematic I had just drawn.
Spent a day on how Bandi looks. Steady flat light → breathing → solar flares → a boiling surface → and then smoothed it all back down.
Bandi started here. Gating, avatar, eyes, triggers, speech bubbles, an action bus, and a prompt cache — all in a single day.
M1–Via–M2 stacking with PECVD ILD and overlay misregistration physics. Free stacking beyond three layers, and a via that can fail to open.
Added two deeper theory slides to every topic deck, then a three-slide learning loop, then unified the level system into one lens.
Started storing the research behind the physics — a deep-research archive wired into Westworld, so a number can be traced back to where it came from.
Sometimes the loop never reaches a clean pass. Instead of hiding that, the document gets a banner saying so.
Generate, critique, revise — with a regression gate that rolls back when the fix is worse than the mistake.
Rebuilt SEM as a tilt-capable 3D engine — full rotation, normal shading, material texture, honest magnification. And 18 images to prove it.
The sidewall angle was being computed twice — once by the etch, once by the renderer. They disagreed. Made the etch the only author.
Added memory/CPU industry timeline simulators, and started rebuilding topic modules as a 3-tier interactive textbook plus presentation decks.
Built a control room for the fab and called it Westworld. Read-only on purpose — you watch, you do not reach in.
Replaced branch-and-threshold behaviour with continuous physics. Process quality now flows into a deficit, and the deficit into the numbers.
Added contamination and process-failure conditions for each tool. The busiest month in this repository — 451 commits.
Google refused to index the module pages and reported one error. Behind that single symptom sat several separate causes — each of them documented somewhere, none of them visible from the browser.
Subtitles drifted ahead of the audio. Fixed it by baking subtitle and audio into each cut before concatenation, instead of laying them over the timeline.
Redesigned every tool to look like the real bench — drag the sample in, read the sticky note on the side, open the manual only if you need it.
The guided walkthrough moves the mouse, clicks buttons and narrates. Getting those three to agree took a month of step-level corrections.
Pulled the etch-rate coefficients back toward published ranges, and put a disclaimer on the screen at the same time.
After SEM, RGA and OES, added an XPS photoelectron spectroscopy simulator to the analysis equipment tutorials.
Pivoted EduAtlas into an academy directory. Seeded 30 in one region, built the registration form and a premium plan. Then stopped.
Rewrote the project brief: GEO engine → AI education knowledge base. Cut categories from six to four on the same day.
Built Beer-Lambert, Planck constant, and ideal gas law simulators in a single day, plus the hub page to hold them.
Stopped making content for an imagined audience and pointed the generator at my own services — semiconductor promo, LittleScience class, blog posts.
The lobby showed eight machines as isometric drawings. Only the ICP etcher actually ran. In March the second one finally became real — ALD.
Scrapped the old landing pages and started EduAtlas. The idea: a document is a record with fields, not an essay.
Came back after a month and nothing ran — the model version in the code no longer existed. Fixed it, then rebuilt the whole narration mode.
The busiest month this repository ever had. Content, Naver SEO, AI crawler access, Firebase Auth, load-time work — all at once.
Turned the equipment picker into a 3D carousel, and drew the ICP etcher as the actual machine — cabinet, chamber, RF, LCD monitor.
Made the PR mask a consumable that carries across steps. Calibrated etch rates to literature and made each material demand its gas.
This repository began as a video generation pipeline. Two weeks later: "Remove all previous project files for new project."
Cropping the black bars kept deleting the audio. Twenty-five commits in three days, and the crop values went hardcoded → auto → hardcoded → auto.
Restructured the course into 3-day, 7-day and 14-day tracks. The same simulators, cut three different ways.
The model kept rewriting the subtitles instead of correcting them. Four prompt fixes later I deleted the step — and brought it back three days after.
Added auto-trim when the conversation hit the token limit. Killed it the next day and made the assistant stop talking instead.
A research plan that sounds good and a research plan you can actually run are different things. Wrote the rules that tell them apart.
The assistant was making up ISEF project names and researchers. Banned it in the prompt and pointed the model at the real database instead.
Pivoted the content from senior lifestyle to 야담 / 전설 / 민담. Added three narrative modes and modularized the camera direction system.
Added valve interlocks, backflow, and five emergency scenarios to the vacuum simulator. The point is what happens when it goes wrong.
The hardest problem of the whole project: keeping one character looking like the same person across every scene. Three approaches in one month.
Added interconnect, inspection and packaging simulators, plus a comprehensive assessment simulator.
Rebuilt the blog around Korean high school science paper coaching. Five days later, hid the posts from the old categories.
The script marks where the arrow should point. Finding a symbol that both the model and the OCR could handle took four tries.
Rebuilt around five genres, added a TTS model router that picks the engine per scene, and pushed a "silent ending" philosophy into every template.
Text animation is easy in a browser and hard in FFmpeg. So I built the scene as an HTML page and recorded it with a headless browser.
Eight commits in one day on a single problem: the highlight has to appear on the right word at the right second.
Preview score and final score disagreed on the same answer. Fixed it by validating that the interpretation matched before reusing.
Cut the RTA temperature response to a seventh, then settled at a quarter. Real process time does not fit a class period.
Started the video factory. Type a topic, get a finished YouTube video: script, voice, images, stock B-roll, rendering.
Converted the site from client-side rendering to static generation to get through AdSense review. Four months later, did it again.
Turned on Firestore so a teacher could hand out an assignment code and a student could open it on any device. Then the browser fought back.
Split the single file into modules, hit build errors, and restored the single file. Twenty-six commits in one day.
An English comprehension tool for teachers. Paste a passage, it picks fifteen sentences and issues an assignment code. Seven commits on day one.
Started a pipeline that turns a slide deck into a narrated lesson video. The split was decided before any code: humans make the diagrams, scripts do the rest.
Five days of work on one file, and every commit message says the same thing. The record of what I did is gone.
Started the teaching simulator: a portal, vacuum, cleaning, and a registry that picks up new modules on its own.
The ISEF competition database sat in an Excel file the browser refused to read. I renamed and moved it four times over six days.
Started a research-writing tool for high school science. The skeleton and the deploy pipeline both went in on the first day.
Before the video factory, this repository held something else entirely: a PWA with astrology charts and AI life advice. 27 commits, then a reset.
index, about, contact, privacy, terms, sitemap, robots, Firebase Hosting. All of it in a single day, thirty commits.
Three days, twenty-three commits, almost all of them titled "Update purchase.html". The page no longer exists.
Added etch, ion implantation and oxidation — completing coverage of the eight core unit processes. Built 3/7/14-day guided courses on top.
The beginning. A vacuum simulator as the first module, then wafer cleaning, deposition, plasma I/II and lithography in sequence.