A Motorcycle in Paju, a Restaurant in Tokyo,
and a Voice at a Bus Stop

April 13, 2026, a Day When AI Entered Individual Hands

1. A Motorcycle in Paju, a Tokyo Back Alley: Tools Made by People Who Do Not Know Code

An intersection somewhere in downtown Paju. A man starts his motorcycle and places his smartphone in the holder. On the screen is an app he made himself. Its name is 'Asajeong'. Across Paju, about 4,000 restaurant addresses have been converted into coordinates, and with one button the navigation turns on with the optimal route. The man designed the app with people in B2B field sales in mind, such as food ingredient delivery or fixtures sales. He said he also rode around on a motorcycle himself. The reaction that followed was, "The efficiency is incredible."

One line he added is the core of this story: "I really do not know even one line of code." A person whose entire experience was typing a few terminal commands called the public data portal API, scraped store information, organized it in Excel, added route optimization logic, built it as a mobile app, and put it on his motorcycle holder.

On the same day in Tokyo, someone else was doing something similar. A developer named exitkibo posted in a thread that he was making a "value-for-money restaurant map for travelers in Japan." The number of shops collected so far is 1,000. Yet he sighed that this is only 1/23 of Tokyo. Nationwide expansion gives him a headache, but he is confident that "once it is complete, choosing restaurants while traveling will become dramatically easier." The target is not tourists but local Japanese users. He added that the model is "local places like the ones Goro from Solitary Gourmet visits."

The two people do not know each other. They are in different countries and use different tools. But they are pushing the same phenomenon forward: vibe coding. This is a concept that platforms such as Knack, Replit, Lovable, Bolt, and Cursor all put forward in early 2026. It means generating full-stack apps with natural-language prompts and finishing deployment inside the browser. Replit's Agent 4 can design and run an entire project by itself, and Vercel's v0 generates production-ready React code based on Tailwind and shadcn/ui. In TechRadar's February 2026 ranking, v0 took first place.

What path would be realistic if implemented in Korea? The starting point is the public data portal, data.go.kr. The "Small Enterprise and Market Service Store (Commercial District) Information API" is public, and with data sourced from the National Tax Service and card companies, it provides business names, addresses, categories, longitude, and latitude. Add the National Tax Service business registration validity API, and you can filter out closed or suspended businesses in real time. The call limit is 1 million calls per day. One city is easy enough. Organize the raw data you scraped with Python pandas, verify coordinates with the Naver or Kakao Maps API, and add 2-opt, an approximation algorithm for the traveling salesman problem (TSP), for route optimization. That is the data layer.

Next comes the app layer. Ask Claude Code or Cursor, "Take the Python script and wrap it as a React Native app." A lighter path is to wrap it with a no-code tool such as FlutterFlow or Glide and deploy it as a PWA, Progressive Web App. If touch use on a motorcycle holder is enough, a PWA is easier to maintain. After deployment, store only user visit records in Supabase or Firebase, and next week you can ask AI in one line for an improvement such as "exclude restaurants visited yesterday."

The real turning point is not technology but psychology. For some people, the sentence "I built an app without knowing a single line of code" is not modesty but liberation. A salesperson builds a tool for their own route, a teacher for their own class, a lawyer for their own case management, and a restaurant owner for their own inventory. Instead of paying a monthly subscription for general SaaS, they make a tool that fits their own hand with a one-day prompt. Paju and Tokyo, two maps, tell the same story today. An app is no longer only a product made by a company. It is a well-worn tool a person uses and discards.

2. The Agent Sits at the Desk: The Arrival of Cowork and Epitaxy

On April 13, 2026, a 'Epitaxy' environment reportedly being tested internally by Anthropic was spotted by the tech outlet TestingCatalog. The report says it is scheduled for release next week. A new UI is being added to the Claude Code desktop app, drawing on the design language of the Cowork layout. Inside one window, three panels, Plan, Tasks, and Diff, sit side by side, and users can open multiple repositories at the same time. On top of that comes a new concept called 'Coordinator Mode.' Instead of Claude implementing directly, it plays only the role of orchestrator, while the actual work is distributed in parallel to subagents.

Why is this an event? Because the old constraint of a single context window is starting to break. Over the past year, the wall Claude Code users repeatedly ran into was the moment when "the context filled up and the conversation had to be restarted." An experimental feature called Agent Teams, introduced after v2.1.32, already existed. A team leader session managed a shared task list while each team member session ran in its own context window. Epitaxy is the step that lifts this from an experimental feature into the default UI.

Meanwhile, Cowork had already changed how many power users work. A user named hamshrew gives one example. Even on sites such as Reddit that block bot access, Cowork opens a browser like a person, searches, takes screenshots, and gathers content. Because it is actual user simulation rather than API calls, it works on most sites. According to leaked documents published on DeepWiki, Cowork runs in a sandboxed Linux VM environment based on Claude Opus 4.6, with 13 core tools and many MCP integrations. It already includes a hierarchical structure in which a parent agent creates special-purpose subagents.

What happens if you put this on top of daily work? Take the AI morning briefing posted every morning on kimkj.com. Until now, a person went through overseas media, selected articles, organized the main points, converted them into KBoard HTML, and uploaded them to the board. In an Epitaxy environment, the flow changes like this. Give the leader agent one sentence: "Prepare today's AI morning briefing." The leader creates a plan and displays it in the Plan panel. Task 1 is assigned to subagent A: "Go through Reuters, Axios, Nikkei, and the People's Daily AI section and collect 30 AI policy and technology articles reported in the last 24 hours." Task 2 goes to subagent B: "From the collected articles, select and summarize five that may affect Korean law and policy." Task 3 goes to subagent C: "Convert the result into compressed HTML for kimkj.com according to the web-html-writer skill rules." The three agents run in parallel in their own contexts, and the leader gathers the results and shows them to the user at once.

This structure has two major implications. One is speed. A pipeline that used to take 45 minutes is reduced to 15 minutes. This is not only because of the mathematical advantage of parallel processing. Each agent focuses only on its own role, so prompt waste disappears. The other implication is auditability. When the Plan, Tasks, and Diff panels are separated, a record remains of "when which agent changed what." In legal work, this recordability is decisive. If AI summarizes client case documents, you must be able to trace through Diff which sentence in which source produced that summary. Epitaxy is solving this requirement at the UI level.

There is also a point to watch. A separate system called Conway has begun to be mentioned together with Epitaxy. According to ChatForest's analysis, Conway appears to be an always-on agent runtime, while Epitaxy is a management and operations layer on top of it. It can be compared to the relationship between Docker and Kubernetes. If this structure hardens, personal developers and companies will move from single sessions to persistent agents. A digital employee that runs on your desk 24 hours a day and works while you sleep. Whether that employee does only good work or strange work is a separate question.

3. A New Genre Called Skills: Prompts Circulate on GitHub

The repository name of K-Dense-AI introduced by an account called vibecoing.school was originally scientific-agent-skills. Now the mirror name claude-scientific-skills is also used. Looking at updates over the past week makes the story clearer. The number has grown from 133 to 170, and then again to 250+ skills integrated with scientific databases. Cancer genomics, drug-target binding, molecular dynamics, RNA velocity analysis, geospatial science, time-series forecasting, FRED economic data. This is not one field. It is the repackaging of an entire research workbench for agents.

The key is that this repository follows an open convention called the 'Agent Skills standard.' Not only Claude Code, but also Cursor, Codex, Gemini CLI, and OpenClaw read the same SKILL.md format. A skill written once runs in multiple AI tools. Installation has become surprisingly easy too. In Terminal, npx skills add K-Dense-AI/scientific-agent-skills one line is enough. It automatically detects directories and places skills in .claude/, .cursor/, and .gemini/ folders.

Occupying a completely different position in the same ecosystem is Paul Bakaus's (@pbakaus) Impeccable. He is a developer known as a former Google Developer Advocate and creator of jQuery UI. Within days of launch, it surpassed 15,000 GitHub stars, and claudemarketplaces now counts it at more than 16,000. It is one of the standout design skills in the Claude Code ecosystem.

The target Impeccable attacks is the phenomenon called 'AI slop.' Frontend code generated by an LLM works, but when you look at it, it smells like AI. Purple gradients, colored stripes on the left side of cards, shadows that are only rounded, the Inter font, old hex values rather than OKLCH. A DartLab dashboard case posted by a user named cowboy76 shows this exactly. The audit score was 6/20. Reasons: repeated use of border-left: 3px solid blue in 14 places, 0 aria-labels, 0 focus rings, and a 756KB bundle. Impeccable breaks these problems into 18 sub-skills and 17 slash commands and processes them sequentially.

The command chain flows like this. /audit runs a five-dimensional diagnosis, /distill removes visual complexity, /polish cleans up token consistency, /adapt handles accessibility and responsiveness, /optimize puts the bundle on a diet, and /typeset replaces fonts. After the audit, cowboy76's score rose 16/20, and the bundle was recorded as going from gzip 204KB to 74KB, 63%smaller. That means the initial 756KB build dropped below the 500KB warning line.

The common thread between the two repositories is this: agent skills are becoming a new 'software distribution unit.' In the past, the units uploaded to GitHub were libraries: Python packages, npm modules, Docker images. Now prompts are uploaded. Role definitions, absolute prohibitions, execution commands, and verification patterns circulate in one SKILL.md file.

From a lawyer's point of view, this is an extension of work already being done. Skills built over the past few months, such as agent-design-prompt, ai-briefing-html-design, column-writer, fact-checker, and transcript-cleanup, all belong to the same genre. Open them to the entire public and they become K-Dense-AI. Focus on one field, design, and they become Impeccable.

The practical implementation path has three steps. First, bundle repeated internal work into a single SKILL.md file: role definition, input format, output format, mandatory rules, examples. Then upload the file to a public GitHub repository and make it installable with an npx skills add flow, like impeccable or scientific-agent-skills. The last step is decisive. Provide audit commands such as /audit so the skill can grade for itself whether the output followed the principles. That is what turned Impeccable from a rulebook into an execution tool. AI legal textbooks, Electronic Times columns, kimkj.com morning briefings. The day when all three are reassembled into their own SKILL.md files and uploaded to GitHub does not feel far away. A skill is lighter than a book, cheaper to share than a lecture, and, decisively, it executes.

4. The Road Down from KakaoTalk: XChat and the Redesign of Data Sovereignty

April 17. Four days later. Elon Musk's XChat officially arrives on the Apple App Store. Pre-orders are already open, and it can be downloaded on iOS and iPadOS. IBTimes and Gizchina reported this within the past 19 hours. The "end-to-end encrypted messenger" that Musk has previewed several times since June 2025 is finally reaching the stage of an independent app launch.

The feature list collides head-on with competing products: end-to-end encryption, no phone number required, disappearing messages, screenshot blocking, message editing, two-way deletion, group chats, file sharing, audio and video calls, and the ability to call Grok AI directly inside the chat window. Musk calls this "Bitcoin-style P2P encryption." In infrastructure language, it is an approach different from the one Signal uses. Messages are not put on a blockchain, but the design creates keys with Elliptic Curve Cryptography and defaults to P2P transmission without going through a central server.

Security researchers are split. Experts cited by CoinDesk and Cryptonews point out that the phrase "Bitcoin-style encryption" is technically confusing. The public-key cryptography used by Bitcoin is for transaction verification, not a protocol that guarantees secret communication between two people. Since the Signal protocol is the proven standard, they question the need to create a new brand. The other side acknowledges Musk's attack point: the suspicion that WhatsApp reads conversation content for ad targeting, and the "ad hook" problem Musk himself mentioned on Joe Rogan's podcast. The logic that this hook is equally open to hackers is structurally correct. If an engine analyzing conversation context to sell ads exists somewhere on a server, a third party that intercepts that engine's input can access the same context. It is also widely known that WhatsApp encrypts message bodies with the Signal protocol but does not encrypt all metadata. Who talked to whom, when, and for how long remains.

Gizchina's point: the App Store privacy card lists collection of location, contacts, search history, and identifiers, contradicting the "zero tracking" marketing phrase.

Apply this to Korea, and the story changes. KakaoTalk has held its status as the national messenger for nearly 20 years, but distrust over its ownership structure and data handling has long surfaced. Where is the data stored? What is the standard for account suspension? If investigative agencies request access, how far does it open? Whether XChat will arrive in Korea as is remains unknown, but alternatives with the same design philosophy are clearly increasing. Signal remains the gold standard, and SimpleX Chat in the open-source camp pushes a structure that allows connection without a phone number.

A personal data sovereignty strategy can be organized into three layers. The first layer is separating messengers by use. Do not put work, family, and sensitive conversations into one platform; split KakaoTalk, Signal, iMessage, and XChat by purpose. The second layer is self-hosting. Put a Matrix-protocol-based Synapse or Dendrite server on a machine you own, such as an Intel Mac Mini Ubuntu server, and move family or internal team conversations there. The third layer is legal control. Using the Personal Information Protection Act's categories of processing consignment and overseas transfer as a standard, create your own ledger of which service stores what, on servers in which country, and in what form.

The real meaning of XChat is not a breakthrough in cryptography. It is marketing that imprints the proposition "no ads means security" on the public. The fact that an argument for separating the advertising revenue model from messaging came from the mouth of a controversial figure like Elon Musk lifts this discussion onto the upstream agenda. Consumers are now belatedly calculating the price they paid for using things for free: time exposed to ads, conversations analyzed, photos used as training data. Looking back at the sum, the word "free" starts to look strangely expensive.

5. The Voice at the Bus Stop: AI in Public Space and Digital Inclusion

On April 13, 2026, Shinil Electronics unveiled an 'AI Human Conversational Kiosk' in partnership with AI software company Saeumsoft. Edaily and Herald Economy reported it that day. It is the result of Anyang City's AI kiosk demonstration project, with Shinil Electronics responsible for hardware design and manufacturing. The project itself was selected under the '2025 AI Challenge Program' hosted by Gyeonggi Province and the Gyeonggi Business & Science Accelerator. The pilot installation sites are four locations.

: Anyang Station, Anyang Jungang Market, Dongan-gu Office, and a bus stop near Indeogwon Station. The kiosk works in a completely different way from existing bus information terminals (BIT). The user asks the digital human on the screen for a destination by voice: "How do I get to Moran Market in Seongnam?" The kiosk combines real-time traffic data and answers by voice and screen with the optimal route, public transportation connections, transfer information, and estimated travel time by transportation mode. It also provides nearby convenience information. If existing terminals were devices that only showed arrival-time numbers, this kiosk is a device that talks. An Anyang City official introduced it as "the nation's first AI digital human bus information system."

Why does this news matter? The answer lies in one phrase: digital inclusion. Generations used to smartphone map apps do not need a kiosk. The problem is the people on the other side. According to material organized by KT Enterprise, Korea's kiosk market grew rapidly after COVID in 2020, but the side excluded by that rapid growth was older people who are less comfortable with digital devices. Turning back at a bus stop because you do not know the route to your destination, or getting on the wrong bus and getting off in an unfamiliar neighborhood. Digitally vulnerable groups were already being pushed little by little out of public spaces.

Shinil Electronics and Saeumsoft's approach is the first experiment to reverse this flow. According to the agreement reported by Easy Economy last September, the hardware is optimized for outdoor environments. Voice recognition works even at noisy bus stops, and it has the durability to withstand summer sunlight and winter vibration. Multilingual support is standard. In addition to Korean, users can converse in English, Chinese, and Japanese, so tourists and foreign residents can use it immediately. The digital human interface lowers the psychological barrier for users afraid of touchscreens because it speaks through the face of a familiar human.

Dissecting this from an implementation perspective reveals three technical layers. At the bottom is the hardware layer: outdoor-durable displays, noise-canceling microphone arrays, protruding speakers, directional cameras. The middle layer is the AI model. It receives questions through STT, speech recognition; an LLM interprets intent; RAG, Retrieval Augmented Generation, brings in real-time traffic data; and TTS, text-to-speech, outputs the answer. The digital human's mouth movements are generated in real time as lip-sync motion matched to phonemes. Commercial solutions such as ESTsoft PERSO.ai are ahead in this area. The top layer is the data layer. It connects the Ministry of Land, Infrastructure and Transport's BIS (Bus Information System) API, public data portal road traffic information, and local government convenience facility databases.

From a legal and policy perspective, there are questions to ask. A conversational kiosk collects the user's voice in real time. Where is this voice stored? How long is it kept? Is it reused as AI model training data? How notice is given and consent is obtained under Article 15 (consent for collection and use) and Article 23 (processing of sensitive information) of the Personal Information Protection Act in an outdoor kiosk environment is a practical issue. When every passerby is a potential data subject, the traditional mechanism of a 'consent button' is almost powerless. One way to solve this is on-device processing. The user's question voice is not sent to a server; it is interpreted, answered, and deleted inside the kiosk. If this design becomes standard, much of the privacy concern will be resolved.

This is the real reason the Anyang pilot project matters. The first button. It is the first experiment placing conversational AI in public space, and the first declaration that rewrites access rights for the digitally vulnerable not as 'consideration' but as 'infrastructure.' A society where you can ask for directions at a bus stop with one spoken sentence even if you have no smartphone, your Korean is weak, or your eyesight cannot read small text. That society is now being piloted in front of Anyang Station.

Closing Scene

AI is now in the hands of individuals and the street. On a motorcycle holder in Paju, in a restaurant map in Tokyo, on a developer's dual monitors, in the messenger inside your pocket, and on the screen at an Anyang Station bus stop. The technology is still large, but the unit that uses it is getting smaller. The age of individuals. More precisely, an age when individuals have begun building their own tools with their own hands.