Shiguang AI Studio

Claude Code for Non-Engineers in Practice・Part 3・About 9 minutes

Teach It Your Habits: How Non-Engineers Write a CLAUDE.md

Contents

    Written with Claude Code 2.1.x, November 2026 (content compiled from official documentation as of November 2026). Claude Code is updating quickly, so screen text, command names, and defaults may change. If what you see differs from this article, please refer to the official documentation.

    What This Article Will Help You Do

    After the first two parts, you've probably noticed something: every time you start a new conversation, it forgets who you are.

    Yesterday you taught it "don't use 'final version' in file names," and today, in a new window, it asks you again. You have to re-explain the naming rules, the report format, which folders not to touch...

    This article solves exactly that: write your habits into a plain text file called CLAUDE.md and place it in your folder. From then on, each time you start Claude Code in that folder, it reads the file at the beginning. Set it up once, and it stays effective long-term.

    By the end, you will:

    • Know what CLAUDE.md is and where to put it
    • Write your first one using a "six-section" structure in 20 minutes
    • Use /memory and /context all to confirm it's actually being read
    • Know what shouldn't go into it, and how to investigate when it doesn't follow your instructions

    What you'll need:

    • Installed and logged into Claude Code following Part 1
    • A folder where you normally work with files (a practice folder works too, such as the Claude練習 folder from Parts 1 and 2)
    • About 30 minutes

    First, understand what CLAUDE.md is and isn't

    What it is: A text file you can write with Notepad (Markdown format, which is essentially plain text with a few # heading symbols). Each time Claude Code starts working, it reads this file as background information. The official description: it's a place to store "what you would otherwise have to explain repeatedly."

    What it isn't: It is not enforced configuration. The official documentation states that Claude treats CLAUDE.md as "background context" and will do its best to follow it, but there is no guarantee of strict compliance. The vaguer or more conflicting the instructions, the more likely they are to be ignored.

    This matters especially for non-engineers:

    Writing "don't delete files" in CLAUDE.md is a reminder, not a lock. The real brake is still the permission mode taught in Part 1, where it asks you before acting. Use both layers together.

    Where it can go (you only need to remember two)

    The official documentation lists several locations. Non-engineers only need to know these two:

    Desired effect File location Description
    Applies only when working in a specific folder The CLAUDE.md inside that folder For example, the Claude練習 folder. If you start Claude Code in this folder (or in one beneath it), it will be read
    Applies no matter which folder you're in ~/.claude/CLAUDE.md ~ on Windows means C:\Users\YourName. Suitable for personal preferences like "who I am" or "always answer in Traditional Chinese"

    Two more details:

    • It reads the CLAUDE.md in the folder where you start and in its parent folders. The contents stack on top of each other rather than overwriting one another.
    • CLAUDE.md files in subfolders are not read at the start. They are loaded only when it begins working on files in that subfolder.

    Recommended approach: start with just one, placed in the folder you work in most often. Add more once you're comfortable with it.

    Step 1: Have it draft one for you (/init)

    If you don't want to start from a blank page, start Claude Code in your working folder and enter:

    /init

    The official documentation says /init automatically generates a starter CLAUDE.md. If the file already exists, it suggests improvements instead of overwriting it.

    Note, however: /init is designed for programming projects. It analyzes the codebase to find build and test commands. If your folder is full of Word, Excel, and image files, the output may be vague or inaccurate. (The author has not tested /init output in a pure document folder; marked "to be verified.") So for non-engineers, I recommend the following instead: write it yourself, or ask Claude to interview you.

    Step 2: Write your first one using the "six-section" structure

    Open Notepad, choose Save As, and name the file CLAUDE.md. Set the save type to "All Files," or it will be saved as CLAUDE.md.txt, which Claude can't read. Put it in your working folder.

    Copy the structure below and replace the contents with your own:

    # 關於我
- 我是【某某公司】的【行政專員】，名字叫【小安】。
- 我不是工程師，請用白話說明你做了什麼，不要丟一大段技術術語。
- 回答一律用繁體中文（台灣用語）。

# 這個資料夾放什麼
- 00 待整理：還沒分類的檔案
- 10 會議：會議紀錄，依年份分資料夾
- 90 封存：已結案，只讀不改

# 檔名規則
- 格式：YYYYMMDD_類型_主題_版本.副檔名，例如 20261105_會議_採購檢討_v1.docx
- 版本從 v1 開始，改版加 1。不要用「最終版」「最終版(2)」。

# 常見工作怎麼做
## 整理檔案
1. 先列出「舊檔名 → 新檔名」給我看，我說「好」才動手。
2. 動手前先把原檔名清單存成 CSV，方便我救回來。

# 不能做的事
- 不刪除任何檔案；要移除的，移到 90 封存。
- 不覆蓋原檔；修改一律另存新版本。
- 不確定的事先問我，一次問一個問題。

# 不要寫進任何檔案的東西
- 密碼、銀行帳號、身分證字號、客戶私人聯絡方式。

    These six sections (who I am / folders / file names / common tasks / things not to do / sensitive data) cover most office situations.

    Follow these four principles when writing

    The official guidance on writing effective instructions, in plain terms:

    1. Be specific enough to check. "File names should be tidy" doesn't work. "The format is YYYYMMDD_type_topic_version, for example……" does. The official example: "Use 2-space indentation" works better than "format code properly."
    2. Keep it short. The official recommendation is to keep each file under 200 lines. Longer files take up its working memory and reduce compliance.
    3. Use headings and bullet points. Content divided into sections is easier to read than a large block of text.
    4. Don't contradict yourself. If a parent folder says "use English file names" and this one says "use Chinese file names," it may pick one at random.

    What to include and what not to include

    Suitable to include Not suitable to include
    Rules needed every time (naming, format, tone) Instructions needed only once (say them directly in the conversation)
    Folder structure and off-limits areas Passwords, account numbers, national ID numbers, client privacy data (these get read into every conversation)
    Corrections for "making the same mistake a second time" A very long operations manual (it gets diluted when too long)

    Another practical rule: the official documentation suggests that if it's a fixed multi-step process, it's better made into a "skill card" (taught in the next part). CLAUDE.md is for "facts and rules you should always know."

    Step 3: Confirm it's actually being read

    Saving the file doesn't guarantee it has been read, so you must verify. After saving, start Claude Code in that folder:

    cd "$env:USERPROFILE\Claude練習"
claude

    Once inside, enter the following two commands (per the official documentation):

    ① View the list:

    /context all

    In the Memory files section of the list, you should see the path to your newly written CLAUDE.md. If it isn't listed there, Claude can't see it.

    ② Open and edit:

    /memory

    This command lists your CLAUDE.md file locations (including user-level and project-level). Select a file to open and edit it in your computer's editor. If a listed file doesn't exist yet, selecting it will create it for you.

    ③ Test it with one question:

    我剛下載了一份採購檢討會議的 Word 檔，請告訴我依照我的規則，它應該叫什麼檔名。先不要改。

    If it answers in a format like YYYYMMDD_會議_採購檢討_v1.docx, it has read the file.

    Note: Screen appearance and section names may vary by version. The /context all and /memory commands above are as documented officially; the author has not checked each version's screens word for word.

    Step 4: When it doesn't follow your instructions, how to investigate

    The official troubleshooting order, in plain terms:

    1. First, run /context all: Is the file listed under Memory files? If not, the file name is wrong (commonly it became CLAUDE.md.txt), or you didn't start Claude Code in that folder.
    2. Instructions too vague? Rewrite them with examples and checkable wording.
    3. Are there conflicting rules? Open each file listed by /memory and check whether two places say different things.
    4. Subfolder CLAUDE.md files don't appear in the list. They're loaded only when needed. This is not a malfunction.
    5. Don't rely solely on CLAUDE.md for things that must always happen. The official documentation notes that actions that must run at specific moments should use hooks (more advanced). A non-engineer alternative is to rely on the permission mode from Part 1, so it asks you before acting.

    Another common question: it also "takes notes" on its own

    Besides the CLAUDE.md you write, the official documentation says Claude Code has "auto memory": it writes its own notes based on your corrections and preferences. These notes are stored in ~/.claude/projects/ on your computer. You can open, edit, or delete them with /memory, or turn the feature off within /memory.

    The distinction is simple: CLAUDE.md is the rules you write; auto memory is the notes it writes. If you want stable, predictable behavior, write the important rules into CLAUDE.md yourself.

    Step 5: Maintaining the file over time

    The official timing guidance is practical, so follow it directly:

    • When it makes the same mistake a second time, write the correction into the file.
    • When you notice you're giving the same correction in different conversations, write it into the file.
    • If you'd rather have Claude write it for you, just say: "Please add this rule to CLAUDE.md."
    • Review it once a month and delete outdated rules. The official docs also provide /doctor prompt-audit, which can have Claude check for outdated or contradictory content (requires a newer version; marked "pending verification").

    Four "don'ts"

    1. Don't put passwords, account numbers, or client data in it. Claude reads it every time, and some people share their folders with others.
    2. Don't write it as a full manual. Beyond 200 lines, the results actually get worse.
    3. Don't assume that writing it locks things down. It's a reminder, not a guarantee. For important actions, still review Claude's plan before you approve it.
    4. Don't write it and forget it. Be sure to verify it once with /context all.

    Want ready-made templates instead?

    Writing your first one from scratch takes the longest. We've prepared templates for three professions: administrative assistant, social media editor, and freelancer. Each one has a finished structure with the "places to fill in" marked, plus a usage guide and rewriting tips. Just replace the 【 】 with your own content.

    👉 View the Product

    (Not buying them is fine too. The "six-part structure" above is enough for you to write a usable first version.)

    Next episode preview

    Part 4: 《Turn Your Weekly Tasks into a Skill Card: An Introduction to Agent Skills》 CLAUDE.md holds "rules you should always know." For the multi-step processes you do every week, the next article shows you how to build a skill card, which you can then invoke with a single sentence.

    Reference sources (all official documentation, accessed November 2026)

    Operating notes: The CLAUDE.md location, size recommendation (200 lines), /context all, /memory, /init, and the descriptions of auto memory are all compiled from the official documentation above. This article has not been tested screen by screen on the author's computer. The output of /init in a folder containing only documents, and /doctor prompt-audit, are "pending verification." The Mac path format (~/.claude/CLAUDE.md) is the same, but the PowerShell commands need to use Terminal's cd ~/Claude練習 instead. The names and companies in the examples are fictional.

    Related Series Reading

    ← Back to article list

    This page was machine-translated from Traditional Chinese. If anything differs, the Traditional Chinese original prevails. Purchases and support are in Traditional Chinese.