Mastering Claude Code

Creating Meeting Summaries and Action Items Tables by Assignee

Transcripts remain after a conference ends, but reading them, summarizing them, and organizing action items by assignee still required manual work. Give Claude Code a single transcript file and one line of instruction, and within minutes a structured meeting record in Word format appears in your folder.

What you need is simple. Whether you use Zoom, Google Meet, Microsoft Teams, Fireflies, Clova Note, or Granola, you can export the transcript as a text file when the meeting ends. Zoom saves it as .vtt or .txt via "Save Transcript," and Teams delivers it as .docx from the meeting details page.

All you need to do is place this file in a transcripts subfolder within your project folder.

Navigate to your project folder in the terminal, then launch Claude Code.

Once the conversation starts, give your instruction in natural language: "Read the file 2026-04-15_exec_meeting.txt from the transcripts folder and save a Word file to the summaries folder that includes a 10-line summary, a list of decisions, and an action items table by assignee. Name the file using the original date and meeting name."

Claude Code internally calls the python-docx library to generate a genuine, editable .docx file. The result contains three elements. First, a summary compressed to ten lines capturing the meeting's nature. Whether a 90-minute meeting or a 20-minute one, the core points emerge at equal density. Second, a separate list of only the decisions reached.

For each agenda item, the format is "What was decided and why?" Third, an action items table with three columns: assignee, task, and deadline. When the transcript contains statements like "I will send it to you by Wednesday," the assignee name and deadline are automatically extracted.

The usefulness of the result is proportional to how specific your prompt is. It is better to provide context alongside your request rather than simply say, "Summarize the meeting." If you write, "This transcript is from an executive meeting on April 15, 2026, with three attendees: the CEO, CTO, and marketing director. Summarize focusing on decisions and omit personality analysis by speaker," the output becomes much more polished.

If you plan to post directly to Slack or Notion, you can specify the format additionally: "For Slack posting, create one more summary as plain text within 300 characters, no markdown. For Notion, use markdown with checkboxes and subheadings."

For recurring work, it is wise to save this process as a standing instruction in CLAUDE.md.

From the next meeting onward, simply drop the file into the transcripts folder and say one line: "Process the new transcript." Even when the session changes, CLAUDE.md rules are loaded each time, so quality remains consistent.

One point deserves attention. Transcripts are sensitive information containing the verbatim remarks of meeting attendees. Before submitting legal consultation meetings, personnel discussions, or client meeting content to an external AI, first verify whether your organization's data policy permits it. In legal work especially, transcripts containing client information require extra caution. Depending on internal rules, you may need to use an API-only path, or follow a process involving anonymization before uploading.

Claude-generated meeting records are drafts. Items where errors are critical,numbers, dates, amounts, legal provisions,must be cross-checked against the original transcript. Especially for specific commitments like "agreed to a percentage increase" or "to deliver by a certain date," you cannot skip reviewing the transcript source.

Automating Downloads Folder Organization

Left unattended, the Downloads folder accumulates hundreds of mixed files: PDF receipts, screenshots, contract drafts, CSVs with jumbled names, and outdated installers all in one place. Finding a file you want again takes minutes. With Claude Code, a single instruction classifies files by type and date, moves old ones to archive, and isolates questionable files,all in minutes.

Before you begin, there is one principle you must follow. File-moving scripts should not actually move anything at first; instead, run a dry-run to see where everything goes. If an important file lands in the wrong folder, recovering it becomes more painful than before cleanup.

In the terminal, prepare your working folder.

Once Claude Code is running, give this instruction: "Create a Python script that organizes the ~/Downloads folder. Route images (.jpg, .png, .gif, .webp) to the Pictures folder, PDFs to Documents/PDFs, videos (.mp4, .mov, .mkv) to Videos, code files (.py, .js, .ts) to Projects/Code, and everything else to Other."

Move files older than six months to Archive, and if you find multiple files with the same name, separate them into a 'DuplicateCheck' folder. Delete nothing, and activate the --dry-run flag by default. Record execution logs in organize_log.txt."

After Claude Code writes the script, run it first with dry-run in the terminal.

Review the output log, and if everything looks correct, proceed with the actual run.

For 186 files, cleanup finishes in roughly 1 minute 30 seconds to 2 minutes. If you instruct the script to read file contents for classification, it takes longer but improves accuracy. For example, if a file named "image_001.jpg" is actually a scanned receipt, the script can read the content and route it to Documents/Receipts instead of blindly sending it to Pictures based on extension alone.

Once you create this task, you need not write it again. Version-control the finished script and run it once a month. You can also register it in crontab to run automatically every Sunday at dawn.

Taking one step further, content-based classification is possible. Gather receipt PDFs in a folder and say, "Read the PDF content of each file and rename it in the format YYYY-MM-DD_vendor_amount. Also keep a changelog.txt recording original and new names." Through OCR and text extraction, bulk renaming happens automatically.

Items where OCR recognition is uncertain get a "[Review Needed]" prefix in the filename, flagging them for human review.

Applied in a law office, this method becomes even more useful. A folder called "Client-Supplied Materials" is created for each case, mixing photos, scanned files, documents, and transcripts. If you establish classification rules tailored to the case type, the folder structure automatically reorganizes each time the client sends materials.

Store categories like "Criminal/Evidence_Photos," "Criminal/Witness_Statements," "Criminal/Client_Notes," "Civil/Contracts," and "Civil/Transaction_Records" in CLAUDE.md.

Stick to one governing principle: file-manipulation scripts should always default to dry-run mode, and actual execution should require an explicit flag. Each time you ask Claude Code to write a script, include this rule in your prompt to minimize mistakes. Automate a 5-minute repeating task with 2 hours of work, and you save 28 hours every year. That is compound interest in time.

Email Classification and Reply Draft Generation

Inbox domination is familiar to everyone. When 200 or 300 emails pile up in a day, important messages vanish among advertisements and automated alerts. Connect Gmail to Claude Code, and within 30 seconds each morning you get a summary sorted into three tiers: "Needs reply," "Worth a look," and "Noise." For important emails, draft replies in your own voice are even saved to drafts.

Two things matter most. First, never auto-send. Second, teach it your voice.

There are several ways to connect Gmail to Claude Code, but using the Gmail MCP connection is most reliable. Create OAuth credentials in Google Cloud Console and register them with Claude Code.

On first run, authenticate your Google account in the browser and the connection is complete. If you have multiple accounts (personal, work, newsletters), register each with a different name. By distinguishing names like gmail-personal and gmail-business, you can give instructions like "Sort only work emails."

First, establish a profile of your voice. Have Claude Code read the 20 most recent messages from your Sent folder and extract the characteristics of your tone.

"Read the 20 most recent emails from my gmail-personal Sent folder and analyze my voice characteristics. Document sentence length, greeting patterns, frequent conjunctions, formality level, and sign-off style, then save to a file called my_email_tone.md."

This .md file becomes the standard for future reply drafts. From then on, give this instruction: "Scan emails received in the last 12 hours and sort them into three tiers: (1) Needs reply, (2) Worth a look, (3) Noise. For the 'needs reply' items, create draft replies applying the tone from my_email_tone.md and save them to Gmail drafts only. Never send them; keep them as drafts."

The VIP list is (CEO, key partner CEOs, corporate card administrator). Prioritize VIP emails to the top."

Here is how results come back. Among 918 unread emails, 7 surface as "needs reply," each tagged with a draft number. Say next, "Make draft 3 a little gentler," and it revises instantly.

To run this workflow automatically each morning, you can set a scheduled hook.

Now, typing "/morning-email" each morning runs the routine. Your only tasks are reading, editing, and sending.

This approach is especially useful for legal work. Hundreds of email threads accumulate per case, exchanged with clients, and every morning it takes thirty minutes just to figure out what progress was made on which case. Tell Claude Code: "Group emails by case using case-number labels, and summarize what happened in each case over the last twenty-four hours in a single paragraph.

Organize and separate the new materials provided by the client, the other party's communication, and the follow-up actions you promised," and within seconds a brief for each case appears.

There is one principle that must never be skipped. Never send an email draft written by AI as-is. A letter to a friend, a proposal to a business partner, a reply to a client,a single word's nuance changes the outcome. You can trust code to AI, but email must be read and revised by a person at the end. An email with errors cannot be unsent once it goes out.

One word wrong in a legal advice email, and the client will act based on that message. Create drafts quickly, but never once violate the principle that a person must hit send. In legal practice, this principle is especially critical.

Batch File Renaming by Reading Content

Receipts, contracts, and scanned tax invoices often have structured content but chaotic file names. To find a specific transaction among names like IMG_2341.jpg, scan_0019.pdf, or untitled-3.pdf, you must open each file one by one. Claude Code reads inside the file directly and renames it according to consistent rules.

The sequence is always the same. First, create a backup. Second, view the results with a dry-run first. Third, execute the actual rename. Fourth, keep a change log.

Prepare your working folder from the terminal.

You instruct Claude Code: "Read the contents of every receipt PDF and JPG file in this folder and extract the merchant name, transaction date, and amount. Rename the files to the format '2026-MM-DD_merchant_amount.extension.' If the merchant name contains spaces, replace them with underscores. For files where extraction is unclear, add '[review-needed]' to the front of the name and record the reason in rename_log.csv. First run it with --dry-run to show me how they will change."

You verify Claude's script with a dry-run. The expected changes print in table form.

Original filename → New filename IMG_2341.jpg → 2026-01-15_Starbucks_4800.jpg scan_0019.pdf → 2026-01-22_Seoul_Garden_87000.pdf untitled-3.pdf → [review-needed]_untitled-3.pdf (merchant name unclear)

If it looks right, you execute it for real.

The change log records the original filename, new filename, extracted data, and confidence score. With just this one log file, you can later trace exactly which files were changed and how if a problem arises.

In law offices, this approach is especially valuable. Thousands of PDFs accumulate in case records, yet most scanned files have meaningless names.

When you instruct it to "extract the case number, document type (complaint/answer/prehearing statement/evidence), and drafting date from each file, rename it to the format 'case-number_document-type_date.pdf,' and move it to a subfolder by case number," the sorting work that would take days finishes in hours. Files where the case number cannot be extracted go to an "unclassified" folder.

It is always safer to instruct it to mark files with unclear classification with the "[review-needed]" prefix. If you accept AI's guesses as-is ("that's close enough"), one or two files in your case records can be classified under the wrong case number. In legal documents, such mistakes are fatal.

The range of uses keeps expanding. For photo folders, you can say "read the EXIF metadata and rename to the format 'date-taken_location.jpg.'" For screen recordings, you can say "analyze the frames and rename to 'date_program_activity.mp4.'" For research papers, you can say "extract author, year, and title from the file and rename to 'author_year_title.pdf.'"

Change only the rules to match how the person who opens the files works, and the same engine organizes any folder.

One crucial principle. File renaming is hard to reverse. If you batch-rename one thousand files and the results are wrong, you can only check the log and undo them one by one. So always follow three rules. Make a backup. Check with a dry-run first. Keep a log. If you do not embed these three rules in your prompt beforehand, you will repeat the same mistake. These three simple lines prevent the kind of accident that wastes an entire weekend.

Converting Receipts and Tax Invoices to Excel

When dozens of scanned receipts and PDF tax invoices pile up, month-end reconciliation becomes a burden. Opening Excel and typing in merchant name, date, amount, and category one by one is simple but tedious and error-prone. Claude Code finishes this task with a single instruction. A real .xlsx file with working formulas is created in your folder.

Preparation is simple. Gather receipt photos and PDFs in one folder and run Claude Code from that folder.

You give the instruction: "Read all receipt images and PDFs in this folder. Extract the merchant name, date, amount, and payment category (meals/transportation/office supplies/entertainment/other). Save to an Excel file sorted by date, with category totals as SUMIF formulas. Mark items where OCR is unclear with 'review-needed' in the remarks column. Name the file receipts_2026April.xlsx."

Within minutes an Excel file appears in your folder. Open it and you see a table with six columns: date, merchant name, amount, category, payment method, remarks. Below the table are category totals as formulas. Change any number in Excel and the totals update automatically.

The key is that the Excel file has working formulas. Claude Code calls the openpyxl library internally and writes not just cell values but also formulas, conditional formatting, and pivot tables to the file. Open it and you see formulas like =SUMIF(D:D,"meals",C:C) actually present.

If settlement categories differ by company, save your category definitions in CLAUDE.md.

From next month on, you just drop files in the folder and say "make this month's receipts into an Excel spreadsheet."

Tax invoices can go one step further. Gather electronic tax invoice XML files from the tax authority in a folder and instruct: "Parse the XML and organize supplier, recipient, item, supply value, tax amount, and total in Excel. Separate sales and purchases into different sheets. Summarize monthly VAT totals in a separate sheet," and your VAT filing basis is ready immediately.

Law offices also use this approach to manage legal fees. When engagement contracts, payment confirmations, and travel expense receipts mix together for each client, settling accounts after a case ends becomes complicated. You instruct: "Read the receipts and deposit confirmations in folders by case number, and make an Excel file with columns separating billable legal fees from reimbursable expenses.

Calculate the total expenses, total fees, and outstanding balance for each case," and a settlement statement ready to send to the client is done.

Some receipts always have poor OCR recognition. Blurry photos, creased paper, ad copy mixed in on the back. AI marks these files as "review-needed," but a person must scan the remarks column at the end and fix them manually. In corporate accounting, every amount must be accurate, and in tax audits, original receipts must match the ledger.

The convenience of automating numbers does not reduce your responsibility to validate them. Whether it takes thirty minutes or thirty seconds to build the Excel file, confirming that the numbers are correct falls to you.

Merging Multiple Documents into One Report

When you wrap up a case or project and open the folder, the file formats are all over the place. PDF research reports, text memos, markdown notes, CSV data, Excel sheets, Word drafts. To merge these into one Word document means opening, copying, pasting, and formatting,tedious work that eats half a day. Claude Code reads the content regardless of file format and assembles it into a single report.

The principle is simple. Claude Code chooses appropriate libraries like pypdf, python-docx, openpyxl, and pandas based on the situation, reads each file, converts content with different formats into uniform text, and assembles the final document. Your only job is to give one instruction: "Here is where the source files are, and here is how I want the output to look."

Prepare your working folder.

Run Claude Code and give the instruction.

"Read all files in this folder and assemble a single case-closing report in Word. Structure it as (1) case overview, (2) chronological timeline, (3) summary of related party statements, (4) evidence list, (5) financial analysis, (6) conclusion and follow-up actions. Standardize all dates to YYYY-MM-DD format and all amounts to Korean won (₩). Convert CSV tables to Word tables.

Add the source filename as a footnote at the end of each section. Save the file as Case_2026_042_Final_Report.docx."

Claude Code internally goes through three stages. First, extraction. It opens each file with the appropriate library and converts it to text. Second, cleaning. It standardizes date notation, currency units, and name format. If the same person appears as "Hong Gildong" and "Hong Gildong," it consolidates to one name. Third, assembly. It rearranges content to match the instructed section structure and converts CSV tables to actual Word tables for insertion.

With many files, Claude Code runs auxiliary agents in parallel. While one reads PDFs, another parses Excel files, and yet another summarizes transcripts. Even a folder with fifty mixed files is organized in minutes.

In legal work, this approach is particularly valuable. When a case runs for months, materials from the client, answers from the other party, witness statements, case precedents, and bank records pile up file by file. When preparing court briefs, reviewing all these materials and organizing the facts takes the longest. You instruct: "Read all materials in this folder and create a chronological factual summary table.

If you instruct 'Be sure to include the source file name and page for each fact. If there are conflicting statements, mark them in a separate section,' you can have the skeleton of a draft memorandum in under an hour.

One thing to keep in mind. The report assembled by Claude is a draft. In a legal document, if one fact is wrong, the entire logic collapses. Based on the source file names and page numbers attached to the end of each section, important facts must always be verified by opening the original documents again. In particular, figures, dates, party names, and contract clause numbers cannot be skipped during the process of checking them once more by eye.

If you have a recurring report format, save a template in CLAUDE.md.

From the next case onward, if you just say 'Make a final report using the materials in the case folder,' a report in the same format will be generated. When the format is consistent, the reader doesn't need to re-analyze the structure for each case, and document management within the office gains consistency.

Automated Weekly Reports: Integrating Notion, Slack, and Google Drive

When you sit down at your desk Monday morning to write the weekly report, the first thing is gathering everything that happened over the past week. You open the Notion project board and count completed issues, scroll back through Slack to find decisions, and check newly uploaded documents on Google Drive. Collecting materials takes an hour, organizing them takes another hour. Claude Code connects all three tools simultaneously and reduces this process to minutes.

First, you connect three connectors. Each of Notion, Slack, and Google Drive provides an official MCP connection server.

Initial authentication for each connector is a single procedure where an OAuth window opens in your browser and you select your account and access scope. With Notion, you can grant access only to specific workspaces and pages, and with Slack, you can restrict access to only necessary channels. For Notion pages containing client information or Slack channels reserved for the legal team, it is safer to deliberately narrow the scope when connecting.

Once the connection is complete, you give instructions in a Claude Code session.

Create a weekly report skill that runs every Friday at 5 p.m.

(Treat messages with an emoji as decisions).

Claude Code calls all three tools simultaneously. It applies database filters in Notion to retrieve only completed items, searches messages by channel in Slack to extract only those with emoji reactions. On Drive, it filters by the modifiedTime field to get only recently changed files. It assembles all these results into a single report and outputs it as a Word file.

You can take this further by scheduling it. Using Claude Code Routines, which launched on April 14, 2026, it runs automatically every Friday at 5 p.m.

Run weekly report routine every Friday at 5 p.m.

Because Routines runs on Anthropic's cloud, it continues running even when you close your laptop or are on a business trip. The Pro plan allows 5 executions per day, Max allows 15, and Team and Enterprise allow 25. If you want to run it only on your Mac, using Desktop's Scheduled Tasks is simpler.

In this case, there is a constraint that the computer must be turned on and the Claude Desktop app must be open, but access to your computer's files is much more flexible.

There are two cautions. First, for sensitive data, you must define the connection scope carefully. Content such as client names, case details, and contract amounts flows through connected services to AI servers, so it is better to use them only when the retainer agreement permits AI tool use, or to create and operate a separate workspace that has gone through anonymization processing.

Second, it is better not to send the draft report directly to your supervisor as is. Sometimes Notion database filters are set incorrectly and completed items get omitted. Have a person review it once at the end, point out 'this item was left out,' and immediately correct and regenerate it. Automation is not a replacement for human judgment but a tool that creates time for people to make judgments.

Competitor News Monitoring and Briefing

Information about the competitive landscape in the legal market, new precedents, and disclosures from companies of interest must be followed daily to be useful. But reading across ten news sites and filtering out what is important consumes an hour or two each day. Claude Code has web search as a tool, so you can say once, 'Organize the new news about these companies today,' and it automatically searches, summarizes, and saves it to a folder.

The key to this task is specifying the search scope and summary format concretely.

Create a competitor monitoring routine that runs every morning at 7 a.m. The target companies are five law firms: Kim & Chang, Gwangjang, Sejong, Yulchon, and Pacific, and the items to check are (1) newly hired partner lawyers, (2) new practice group launches, (3) major engagement announcements, and (4) media interviews and column contributions. Retrieve only content published in the last 24 hours and focus on Korean-language news.

Save the results as a '~/intel/YYYY-MM-DD_law_market_brief.md' file, and be sure to include the source URL at the end of each item. Classify by importance level as high/medium/low and gather only high-priority items at the top.

Claude uses the web search tool to search by combining each company name with keywords. It creates queries that include dates, like "Kim & Chang partner hiring 2026" or "Gwangjang new group launch," reads the top results, and extracts the essentials. The results are saved as a markdown file and can be opened directly in Obsidian or Notion.

The same pattern applies not only to company news but also to precedent and legislation monitoring.

Create a new precedent briefing that runs every Monday morning at 8 a.m. Areas of interest are (1) AI and data-related precedents, (2) Personal Information Protection Act violations, and (3) startup investment contract disputes. Search for precedents issued and released last week on domestic court sites (law.go.kr, scourt.go.kr) and major law firm blogs.

For each precedent, organize the case number, issues, and holdings, and add a 2 to 3 sentence explanation of why it matters in practice. Save the results as '~/intel/YYYY-MM-DD_precedent_brief.md' and suggest 3 blog posting ideas in the '~/blog_ideas.md' file.

You can go one step further and not just save the collected information but turn it directly into writing. After the competitor trend summary, if you add, 'Pick 3 items from this information that would work well as blog post topics and write an outline for each,' you get content planning materials along with the briefing each morning.

The same pattern works in personal matters too. Companies in your investment portfolio, changes to the terms of services you use, major papers in technology fields you care about checked daily for a daily summary. Check the stock price-related news and disclosures for Samsung Electronics, SK Hynix, and Naver from the last 24 hours. Place stocks with price changes of ±3% or more at the top.

If you instruct, 'For important disclosures, summarize the full text,' you get foundational materials for individual stock analysis stacking up in a folder every morning.

Web search results have one blind spot. Source reliability varies widely. News retrieved by AI is sometimes a blog summary or an article filtered through two or three layers, and rarely it can be an older article with an incorrect date attached. So you instruct to always keep the original URL as the source in the results file. For important content, a person opens the original once more to verify.

In particular, personnel movement news like 'who moved where' is often confused with a different person of the same name, so it is necessary to develop a habit of viewing photos and career descriptions from the original source together.

Different care is needed when handling market analysis reports or economic indicators. AI is good at stating numbers plausibly, but does not guarantee the accuracy of the numbers themselves. Even if a sentence appears in the briefing saying 'crude oil prices rose 4.2% this week compared to last week,' it could actually be 3.8%.

If numbers form the basis of decision-making, you must go down to the original data sources (Bank of Korea, Statistics Korea, exchange disclosures) once more to verify.

Competitor monitoring has information whose value is time-sensitive. If you see news from yesterday a week later, its usefulness is cut in half. That is why it is important to run this task automatically every day, and if you put it in Routines, it runs in the cloud even when your laptop is closed. By the time you go to the kitchen to make coffee each morning and come back, today's market briefing is sitting in a folder on your desk.

Automated Daily Morning Briefing Generation

How you spend the first 30 minutes of your day determines the density of your entire day. If you open your inbox and start getting pulled around by scattered emails, you will spend the morning just reacting and never reach the work that truly needs your focus. The morning briefing flips this starting point. Before sitting down at your desk, you create a state where today's important items are summarized on a single page.

The structure is this: today's schedule, urgent emails needing replies, follow-up actions promised in last meeting, focus time blocks that should not be interrupted. You need just one markdown file where these four things are visible at a glance.

You create the briefing skill in the terminal.

You create a new task in Claude Code Desktop's Scheduled Tasks, write '/morning Run the skill' in the prompt, and set it to run at 7 a.m. daily. If your laptop is on and the app is open, the briefing file is generated automatically every morning. If you use Routines, it runs in the cloud, so it works regardless of your laptop's state.

When you open the results file, it looks like this. At the top are three items for today's schedule. Next to the 10 a.m. client meeting is a note saying, 'Opinion letter draft promised at last meeting; not yet delivered.' Next to the 2 p.m. consultation call are two lines summarizing the last call and one line of the counterparty company's recent press release. Below that are three emails needing urgent replies.

An internal email from the CEO, a contract inquiry from a key business partner, a court hearing schedule notice from the jurisdiction. Each has a one-sentence explanation of why it is urgent. Below that are four follow-up actions promised at the last meeting. Items like 'materials promised to send by Wednesday,' 'email promised to reply by Friday' are listed alongside their original promised dates.

At the bottom, the note reads: 'Secure one deep work block of at least 90 minutes today.'

This approach holds particular value in legal work. When you manage multiple cases simultaneously in a single day, reorienting yourself to each case's current status consumes considerable time. If you add a 'today's tasks by case' section to your briefing, each case's deadline and next steps refresh automatically and surface at the top. Looking at the case name alone, you can set priorities: 'Case X: filing deadline tomorrow,' 'Case Y: witness examination preparation needed.'

For a briefing to serve its purpose, one thing is essential: it cannot be too long. One page, readable within three minutes. A briefing listing twenty items is not a briefing but simply another to-do list. Build this constraint into your prompt: 'compress to one page, readable within three minutes.' Automation exists to deliver what is necessary quickly, not to generate more.

Reorganizing scattered notes into a logical draft

A writer's folder accumulates scattered notes at any moment. A single line recorded on an iPhone when an idea struck. A sketch scribbled on paper during a meeting, then scanned. Markdown notes typed while listening to a lecture. Quotations copied while reading a book. A paragraph tossed into a KakaoTalk 'notes to self' chat.

These fragments clearly hold connections among themselves, yet assembling them into a finished piece requires time and concentration. Claude Code reads these fragments, groups them by subject, grasps the logical flow, and reconstructs them into a draft.

Start by gathering files into one folder. Format need not be uniform. Place text files, markdown, Word documents, PDFs, and transcriptions of voice memos all in the same folder.

Run Claude Code and give it instruction.

"Read all the files in this folder. Group them by subject, and merge ideas that overlap, but preserve the subtle nuances of meaning in the original text. Then add subheadings by subject and reorder them according to logical flow. Assemble them into a draft manuscript in Korean and save as 'AI_Law_Draft_v1.md.' At the end of each paragraph, note the source file in brackets so I can find the original later."

Keep the author's voice in formal style throughout ('this is,' 'this does'), and refer to project_style.md for the forbidden expression list."

The result is not a finished manuscript. Pieces have been classified, gathered by subject, and reordered in ways that create potential logical connections. The writer then reads through this again, edits the flow, and refines it in their own voice.

This approach matters because it lightens the burden of memory. When a single-line note written six months ago and lecture notes from last month are grouped into the same topic, the writer finally recognizes: 'Ah, these two could connect.' People delight in such connections when chance reveals them, but when hundreds of files accumulate, chance itself no longer occurs. When AI scans the fragments and classifies them, the very possibility of connection unfolds before you.

Writing by lawyers demands special rules. Citation and attribution are the lifeblood of such text.

With such rules in place, you can toss notes into a folder and say only 'make a draft,' and the result emerges while preserving the writer's style. Another writer's folder contains their own rules, and each time AI shapes the result to fit those rules.

One caution applies: AI excels at reorganizing 'logically,' but logical does not always mean good writing. Good writing sometimes breaks logic, begins with a scene, and leaves conclusions suspended. If you use AI's organized draft as-is, the writing becomes smooth but flat.

What the writer must do is receive AI's organized version and then ask: 'Where do I start with a scene?' 'Which section should I deliberately skip?' 'Which sentence should I move to the second-to-last line to create lingering effect?' AI lines up the fragments. The writer scatters the lines to build rhythm. The clearer this distinction between the two tasks, the better the outcome.

Artificial intelligence legal policy specialist Attorney Kim Kyung-jin

AI policy and law expert · Former National Assemblyman · Author of numerous works

If this book lingered with you even for a moment, please support it so the next story can reach the world.

(Voluntary donation account: Nonghyup 302-1096-0948-81 Account holder: Kim Kyung-jin)