Table of Contents
Artificial Intelligence Translates the Language of Animals
Kim Kyung-jin, Attorney at Law
The Story of AI Learning to Listen to Whales, Dolphins, Birds, and Bees
Twelve chapters on how AI listens to dolphins, sperm whales, humpback whales, birds, and bees to find rules in their sounds, and what this technology means for its risks and for animal rights. Written in simple sentences a child can read, with verified sources in every section.
Table of Contents
AI Deciphers Ancient Scripts
Kim Kyung-jin, Attorney at Law
Ancient Records Revived by AI
In twelve chapters, this book explains how AI revives records once unreadable, from burned scrolls and wooden slips buried in mud to broken clay tablets. It covers virtual unrolling at Herculaneum, virtual collation of oracle-bone texts, reading Silla wooden tablets, computational analysis of undeciphered scripts, and multispectral archives, with verified references for each chapter.
Table of Contents
Artificial Intelligence for New Materials Design and Rocket Propulsion Engineering
Kim Kyung-jin, Attorney at Law
AI Potentials, Self-Driving Laboratories, and Physics-Informed Machine Learning (PIML)
Ten chapters on how artificial intelligence is changing new materials and rocket propulsion: atomic simulation, generative models, self-driving labs, high-temperature alloys, metal 3D printing, combustion, cooling design, and engine diagnosis and control. Written without equations, with verified sources in every chapter.
AI Library
The Age of Autonomous Scientific Discovery
Kim Kyung-jin, Attorney at Law
AI Scientists and Self-Driving Labs
This book follows how AI scientists and self-driving labs are changing the way science generates and verifies claims. It covers literature-based discovery, natural-language protocols translated into robot commands, multi-agent research systems, closed-loop laboratories, materials search, the verification gap, chains of evidence, research harnesses, journal ethics, and legal responsibility.
AI Library
A New Era of Life Sciences Opened by Artificial Intelligence
Structural Proteomics, Genomic Foundation Models, Autonomous Laboratories, and Global Governance
Kim Kyung-jin, Attorney at Law
This book is a research volume compiled with artificial intelligence. A human selected the materials and structured the work, while AI models drafted the sentences and cross-checked the facts.
AI Library
The Double Structure of Digital Sovereignty
Europe’s Departure from Palantir and the Chains of American Big Tech
Kim Kyung-jin, Attorney at Law
This is a record of 2026, when European intelligence agencies and defense ministries began removing analytics tools from America’s Palantir. It covers the replacement decisions made by France’s General Directorate for Internal Security (DGSI), Germany’s Federal Office for the Protection of the Constitution (BfV), and the Netherlands Ministry of Defense; the incident in which US export controls severed an ally’s ac…
New English Edition
Artificial Intelligence in Horticulture
Kim Kyung-jin, Attorney at Law
Across five chapters and ten sections, this book examines computer vision for crop diagnosis, harvesting robots and autonomous field systems, smart greenhouses and digital twins, precision irrigation and supply-chain quality control, high-throughput phenotyping, and predictive breeding.
New English Edition
Artificial Intelligence in Food Crop Agriculture
Kim Kyung-jin, Attorney at Law
Across six chapters and eighteen sections, the book examines digital agricultural infrastructure, remote sensing, crop diagnosis, yield forecasting, precision irrigation, genomics, molecular breeding, agricultural robotics, climate-smart agriculture, and global food security.
New English Edition
The Future of Forestry and Agroforestry
Kim Kyung-jin, Attorney at Law
Driven by Artificial Intelligence and Digital Innovation
Across five chapters and fifteen sections, the book follows satellites, drones, LiDAR, digital twins, forest-specific language models, wildfire and pest forecasting, forestry robotics, agroforestry, timber traceability, and forest carbon markets.
New English Edition
Smart Livestock Farming: AI Enters the Barn
Kim Kyung-jin, Attorney at Law
Sensors listen, cameras watch, and artificial intelligence helps farmers decide.
Across five chapters and fifteen sections, the book follows precision livestock farming from animal health and reproduction to robotic milking, virtual fencing, digital twins, methane reduction, welfare, and data ownership.
Table of Contents
Han Dong-hoon, Busan Buk-gu Gap: A Record of the 100 Days Before and After the Election (Mar. 26-Jul. 3, 2026)
Kim Kyung-jin
Table of Contents and 13 sections
From March 26 to July 3, 2026, this record follows the spring after expulsion, the Busan Buk-gu Gap by-election, victory as an independent, and the first bill submitted in the National Assembly.
Table of Contents
Artificial Intelligence and Medicine
Kim Kyung-jin, Attorney at Law
AI in clinical care, hospitals, education, and research
AI in medical imaging, risk prediction, treatment planning, hospital operations, education, and research, with patient safety, privacy, and accountability.
[AI Library] Chapter 7: IT Projects and Software Development
Claude Cowork and Agent Utilization Manual
Chapter 7: IT Projects and Software Development
Kim Kyung-jin
Building software is similar to constructing a house. You draw blueprints, select materials, build the frame, finish the interior, and check for leaks.
The most time-consuming work in this process is not the act of laying bricks, but deciding where and how to lay them and verifying that they are laid correctly. Claude Code and Cowork reduce the burden on developers and planners in the areas of decision-making and verification.
This chapter differs in character from the previous chapters 1-6. Although it deals with code, we have lowered the threshold so that readers without coding experience can follow along in planning areas such as writing requirements documents (PRDs) and creating test case documents. Conversely, in areas requiring development experience such as operating agent teams or code migration, we prioritize conceptual understanding, while execution guidance is aimed at intermediate and advanced readers.
1 Parallel Work Based on Agent Teams
A One Developer Doing the Work of Four
Imagine building a single web application.
You need to build the user-facing interface (frontend), establish a server that processes data behind the scenes (backend), and design a database to store the data. Finally, you must test whether all three work together correctly.
If one person does these four tasks sequentially, the next task cannot begin until the previous one is finished. The frontend cannot retrieve data without a completed backend, and backend code cannot be written without a finalized database structure. This sequential bottleneck is the biggest cause of prolonged development time.
In actual companies, teams solve this problem. Frontend developers, backend developers, database administrators, and QA engineers work simultaneously on their respective areas and communicate and coordinate with one another as needed. Agent Teams reproduce the structure of human teams using AI.
B What Are Agent Teams?
Agent Teams are a function where multiple Claude Code instances (individual AI sessions running simultaneously) each take on independent roles to execute a single project in parallel. They were released as a Research Preview alongside the launch of Claude Opus 4.6 in February 2026.
Previously, there was a concept called sub-agents.
Sub-agents operated by following instructions from the main agent and returning only results. The problem was that sub-agents could not communicate with each other. There was no way for the frontend sub-agent to directly tell the backend sub-agent, "The API address has changed, so fix the code." All messages had to pass through the main agent, and this bottleneck frequently caused code conflicts in complex projects.
Agent Teams changed this structure. Team member agents can exchange messages directly with each other. Through a Shared Task List, they can see who is doing what work, and when one task is complete, they take on the next one themselves. An agent in the Team Lead role coordinates the overall effort, while practical communication happens directly among team members.
[Note] What is an agent instance? It refers to a single Claude Code session. Each instance operates independently with its own context window (the range of information an AI can remember at one time). Claude Opus 4.6 supports a context window of 1 million tokens (approximately 750,000 words) per instance, allowing you to work with an entire large codebase at once.
C How to Use Agent Teams
(1) Basic Use: Enabling Agent Teams and First Run Agent Teams are an experimental feature and must be enabled separately.
① Open Claude Code's settings file (settings.json).
② Add the following line: "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1"
③ Save the settings and restart Claude Code.
④ In the terminal, give the following instruction: "Analyze the README.md file and entire folder structure of this project, then create an agent team with two members,one for security review and one for code quality review,and perform a code review."
⑤ Claude Code assumes the Team Lead role and creates two team member agents. The activity log of each agent is displayed in the terminal.
⑥ Team member agents each analyze the code and share any issues they find with each other through messages. The Team Lead synthesizes the results and generates a final review report.
(2) Practical Example: Splitting Screen Views by Agent with tmux If you install tmux (a terminal multiplexing tool), you can divide each agent's work screen and observe them simultaneously.
① When you run an agent team with tmux installed, the screen divides into multiple panels, with each agent's terminal displayed separately.
② Press Shift plus an arrow key to move to a specific agent's panel, then give instructions to only that agent.
"Frontend agent, change the button color of the login page you just created from blue to navy."
③ The agent receives the instruction, makes the change, and notifies the team members of the modification.
(3) Real-World Application: Building a Blog Platform Project A scenario for building a small to medium-sized blog platform.
① Prepare a requirements document in the project folder. (The method for writing a PRD is covered in Section 2 of this chapter.)
② Give Claude Code the following instruction:
"Read this PRD and create an agent team consisting of a backend API lead, a frontend UI lead, a database schema lead, and a test writing lead. Have the backend agent implement a REST API using Node.js and Fastify, the frontend agent implement the UI with React. Have the database agent finalize the PostgreSQL schema first and share it with the other agents. Have the test agent write unit tests for each API endpoint."
③ Claude forms the team and distributes work. Once the database agent finalizes the schema first, the backend and frontend agents simultaneously write code based on that schema.
④ As work progresses, the Team Lead provides a summary of status. If conflicts arise, the Team Lead judges priorities and resolves them.
D Things to Know When Using Agent Teams
Agent Teams consume a large amount of tokens. Because each team member maintains an independent context window while exchanging messages with others, costs can increase 4 to 7 times compared to having a single agent do the same work. There is no reason to use agent teams for simple tasks like script writing or file organization.
The situations where agent teams prove effective are clear: projects that must simultaneously touch multiple layers (frontend, backend, database, testing), debugging that requires validating different hypotheses at the same time, and large-scale code reviews where parallel exploration creates real value.
One more thing. While code produced by agent teams may look clean individually, it can cause unexpected conflicts within the context of the overall system. A caching structure introduced by one agent for performance may conflict with another agent's
real-time synchronization logic. Agents may write code faster, but the responsibility for judging how that code operates within the overall system still rests with humans.
E If These Problems Occur
(1) If an agent team is created but only one message is sent and the session terminates, verify that the agent team settings have been correctly saved in settings.json. Check that quotation marks and colon placement are exact, then completely close Claude Code and restart it.
(2) If team member agents simultaneously modify the same file, causing code to break, specify file ownership in the prompt. Instructing each agent to modify only files in its assigned folder and read-only access to other agents' files can reduce conflicts.
(3) If token usage runs much higher than expected, reduce team size. For tasks that five people can do with three, adding extra members only increases coordination costs. The Max pricing plan suits agent team work better than the Pro plan.
2 Software Requirements Definition and Architecture Documentation
A Why "Make an app with payment features" fails
The most common failure in asking AI to write code stems not from lack of AI capability but from vague instructions. When you say "Make a project management app with payment features," AI fills in numerous blanks on its own. Is payment by card or bank transfer? Is user authentication email or social login? Is project management organized by team or individual? When the assumptions AI supplies differ from your intent, the result is something nobody wanted.
A Product Requirements Document (PRD) is a document that fills in these blanks ahead of time. It clearly states what will be built, who it is for, what success looks like, and what will not be done this time. PRD quality determines whether a project succeeds or fails.
B What kind of PRD does Claude Code produce?
With Claude Code's Plan Mode, AI focuses entirely on analysis and design without writing a single line of code. Feed it fragmented ideas floating in your head, and AI asks systematic questions to fill the gaps, then organizes the results into a multi-page planning document.
The final deliverable includes the following. Service purpose and success criteria, core user journeys, functional and non-functional requirements (performance, security, availability), database schema (table relationships, data types, constraints), system architecture (tech stack, infrastructure layout, authentication scheme), and test scenarios.
C Let's create a PRD
(1) Basic use: extracting a question list from an idea
1. Open Claude Code terminal and switch to Plan Mode.
2. Enter the following:
"I want to build a secondhand marketplace app. Location-based search, chat, and escrow payment are the essentials. Before I write code, give me ten or more clarification questions to reduce assumptions about my project."
3. Claude presents questions like these: "What distance radius should location-based search use in kilometers?" "Will chat support image uploads?" "What commission rate for escrow payment?" "Is signup email verification or social login?" "Do you need seller authentication?"
4. Answer each question.
5. Claude generates a PRD draft based on all your answers.
(2) Advanced use: including database schema and API specification
1. Review the PRD draft from basic use, then enter additional instructions:
"Design a database schema in Prisma format based on this PRD. Include User, Product, Transaction, Chat, and Review tables, and specify each relationship and index. Then organize core API endpoints in OpenAPI format."
2. Claude generates a Prisma schema file and OpenAPI specification.
3. Open the generated documents and review the content.
(3) Real-world use: Using PRD as the agent team's execution guide. PRD's real value lies in its role as the starting point for coding.
1. Once PRD is complete, reset your Claude Code session. Mixing previous conversation into the context window creates confusion.
2. In the new session, give this instruction:
"Read the PRD.md file, schema.prisma, and openapi.yaml in this folder. Assemble an agent team according to these documents and build the app. Use Fastify for backend, React for frontend, and PostgreSQL for database."
3. The agent team begins work, treating the PRD as the source of truth for execution.
D Why you cannot use AI-generated PRD as-is
AI-generated PRD is a starting point, not a final document. AI excels at creating logically airtight structures but does not fully understand subtle business context differences. Real-world constraints like Korea's escrow payment regulations, PIPA data retention requirements, and UX preferences by customer age must be reviewed by people.
It works well to sequence review this way. The product owner confirms goals and scope. Engineers assess technical feasibility. Security staff verify authentication, encryption, and logging policy. Operations review deployment and monitoring systems. Good design documents do not emerge complete in one pass; they crystallize decisions through review cycles.
E Territory non-developers can master
PRD writing is achievable without coding experience. Explaining your service idea and answering AI's questions belong to business sense. The same process works on Cowork's desktop interface, so readers unfamiliar with terminal should start with Cowork.
F If you run into these problems
(1) If the PRD feels too generic and misaligned with your project, it is because you skipped the clarification step and jumped straight to the PRD. Ask AI to pose questions first, answer them specifically, then have it generate the document. Quality improves.
(2) If the generated database schema is missing required tables, specify the gaps explicitly: "Add Audit Log, Notification, and Attachment tables."
3 Code Migration and Automated Testing
A The invisible debt called technical debt
Technical debt is the invisible cost accumulated by old code. Each time you add a feature, bugs surface in unexpected places. New developers need weeks just to understand the codebase. Applying security patches hits dependency snags. So companies periodically run code migrations,swapping one framework for another.
It is among the tasks development teams dread most. You are not building something new; you are changing internal structure while keeping existing features intact. Users see no visible change, yet time and cost investment is enormous.
B How Claude Code helps with migration
Claude Opus 4.6's one-million-token context window lets you analyze entire codebases of hundreds of thousands of lines at once. Older AI tools had to break code into small chunks, so inter-file dependencies were easy to miss. Now you can build transformation plans while holding the full architecture in view.
Migration typically follows four stages. First, identify the boundaries of your current code. Separate routes, middleware, authentication, error handling, database access, and external API calls. Second, freeze the current API structure in OpenAPI or JSON Schema. Third, reorganize to fit the new framework's structure. Fourth, run the old and new versions side by side and verify behavioral differences.
C Migrating from Express to Fastify
(1) Basic usage: Creating a migration analysis report① Connect your existing Express project folder to Claude Code. ② Give the following instruction.
"Analyze this entire project. Find all routes using the Express app.get and app.post patterns, map the middleware chains, and catalog where database queries execute. Don't modify the code; instead, create a migration analysis report that assesses the risk level of converting to Fastify on a scale of 1 (low) to 5 (high)."
③ Claude scans the project and generates a report analyzing the cross-references among all files and identifying risk factors.
(2) Applied case: Executing the conversion step by step
① Review the analysis report, then begin the conversion starting with modules that have low risk.
"Convert the routes rated 1 to 2 for risk first to the Fastify format. Replace Express's req and res objects with Fastify's request and reply objects, and add JSON Schema-based request validation to each route."
② Claude converts those files. Rather than applying mechanical substitution, it reflects the structural differences between Express's middleware approach and Fastify's plugin approach.
(3) Real-world use: Auto-generating and validating tests
Verifying that migrated code works correctly is just as important as moving the code itself.
① Give the following instruction. ②"Generate unit tests for each converted Fastify route. Include scenarios for normal requests, invalid parameters, authentication failures, missing resources, and server errors. Use Fastify's inject method so tests can run without opening a network port."
② Claude generates test code for each endpoint.
③ Run the generated tests to check if they pass. "Run all the tests you just created."
④ Claude runs the tests; if any fail, it analyzes the cause and fixes the code. This process repeats until all tests pass.
If end-to-end (E2E) testing is needed, integrate Playwright (a browser automation tool). "Using Playwright, create E2E test code that tests the entire user flow of logging in, searching for products, adding items to the cart, and completing checkout."
The Limitations of AI-Generated Tests
AI-generated tests verify the code's current behavior as-is. If the current behavior itself contains a bug, a test that treats that bug as normal can be created. For this reason, human review of AI-generated test code is necessary. The process of confirming whether "the behavior this test validates matches the behavior we intended" cannot be automated.
Logic left by past developers that appears irrational may actually have been a workaround for a bug in a specific operational environment. AI, unaware of this hidden history, can refactor the code away. Determining what must be preserved during migration is a human responsibility.
When These Problems Occur
(1) If a "Cannot find module" error repeats during migration, it means a dependency package is incompatible with the new framework. Instruct Claude: "Analyze the dependency list in package.json and find packages that are incompatible with Fastify."
(2) When all tests pass but errors occur in the live service, it means there are scenarios the tests don't cover. You must reflect real data patterns from the operational environment or conditions like network latency in your tests.
4 UI/UX Web Design and Figma Synchronization
The Old Divide Between Designers and Developers
When a designer carefully crafts a design mockup in Figma (a web-based design tool) and a developer implements it in code, subtle differences emerge. Margins are off by four pixels, font weight is applied as bold instead of semibold, and shadow opacity differs. The process of handing over a design mockup to a developer is called handoff, and this phase has been one of the greatest friction points in product development.
AI is reducing friction in this handoff. It approaches this from three directions.
What Is Figma MCP Integration?
MCP (Model Context Protocol) is a communication standard that allows AI to exchange data with external tools. With Figma providing an MCP server, Claude Code can now directly read and write design information from the Figma canvas. The "Code to Canvas" feature, released on February 17, 2026, enables bidirectional synchronization.
The Figma MCP server doesn't just see the pixels in a screenshot. It reads structural information such as component hierarchy, color codes, font sizes, spacing values, and design tokens (values that define the basic units of design).
Connecting and Using Figma with Claude Code
(1) Basic usage: Setting up the Figma MCP server
① Update the Figma desktop app to the latest version.
② Enter the following command in the Claude Code terminal.
claude mcp add --transport http figma https://mcp.figma.com/mcp
③ Enter the /mcp command in Claude Code to verify that the Figma server is connected.
④ When Figma requests authentication, click "Allow Access."
(2) Applied case: Converting a Figma design to code
① Select the frame (a screen unit) in Figma that you want to convert.
② Paste the Figma link into Claude Code and give the following instruction.
"Analyze this Figma design and convert it to a React component. Use Tailwind CSS and apply responsive breakpoints. Preserve the exact color codes and font sizes defined in the design."
③ Claude reads the design's structural information through the Figma MCP and generates React component code.
④ Run the generated code on a local server and compare it with the design mockup.
(3) Real-world use: Sending code back to Figma in reverse
This feature is the key to achieving bidirectional synchronization. Developers can send the UI they built in code to the Figma canvas, allowing designers to receive it as an editable frame.
① With a local server running in Claude Code, issue the following instruction:
"Start my app's local server, capture the current screen, and send it to a new Figma file."
② Claude opens a browser window and displays the capture toolbar. You can capture the entire page, specific elements, or multiple screens consecutively.
③ Once capture is complete, Claude provides a Figma file link. When you open this link, the captured screen appears as an editable Figma frame.
④ When a designer modifies colors or spacing in Figma, Claude Code can detect those changes and apply them to the code. "Check for changes to this frame in Figma, and apply only the modified parts to the code."
D. Constraints to Keep in Mind
Some features of the Figma MCP server operate only under specific conditions. The Code to Canvas feature is currently supported only in Claude Code, OpenAI Codex, and VS Code. Figma also requires a Dev or Full plan with Organization or Enterprise tier. API calls are subject to rate limits, so converting large numbers of frames at once may slow performance.
While the visual accuracy of UI code generated by AI is improving, it does not automatically meet accessibility standards. Details such as screen reader compatibility, keyboard navigation, and color contrast ratios must be validated separately by humans. Interaction logic such as drag and drop, infinite scroll, and complex form validation shows uneven quality in code produced by AI.
AI can quickly produce the visual appearance of a design. But the responsiveness felt at the user's fingertips, the cadence with which the screen appears when scrolling, and the tone conveyed by error messages remain areas that still require human refinement.
E. Practical Implementation in Three Stages
Trying to implement all features at once will exhaust your team. In the first stage, learn the basic workflow of converting design mockups into code. In the second stage, connect the Figma MCP to supply AI with design context (colors, spacing, component hierarchy). In the third stage, set up Code Connect (which links Figma design components with actual components in your code repository's
tool), so that instead of AI inventing new components, your team uses components it already has. Moving methodically upward is ultimately faster.
F. If You Encounter These Issues
(1) If you have connected the Figma MCP server but the generate_figma_design tool does not appear, a conflict may occur if a previously connected Figma MCP instance remains. Disconnect the existing connection, restart Claude Code, and reconnect. Installing through the Figma plugin is more stable.
claude plugin install figma@claude-plugins-official
(2) If the result of converting from Figma to code differs significantly from the original, working from a screenshot alone produces different results than reading the structure through MCP. Verify that MCP is properly connected using the /mcp command, and explicitly state in the prompt: "Fetch design context from Figma MCP and perform the conversion."
(3) If the Code to Canvas feature does not work, this feature is supported only through the remote MCP server (https://mcp.figma.com/mcp). If you have connected only the local MCP server (the Dev Mode MCP Server in the desktop app), the feature to send code to Figma will not work. Connect the remote server in addition.
The four areas covered in this chapter are independent features, but in actual projects they form a single workflow.
Requirements are defined in a PRD, the agent team implements the code, automated testing validates quality, and Figma sync closes the gap between design and code. AI codes faster than humans. But the decision of what to build and judgment about whether the result aligns with business context remain the responsibility of the person at the screen. Placing our organization's unique judgment and experience atop the structure the machine has built. That is the right way to use these tools.
Kim Kyung-jin
Attorney · Former Member of the National Assembly · AI Policy Researcher
© 2026 Kim Kyung-jin. All rights reserved.