Optimizing Claude AI Context with a Standardized 4-File System Architecture

Claude Pro costs $20 per month, roughly 500,000 VND depending on exchange rates and bank fees[1]. Many people pay that amount just to have Claude write emails or summarize a few paragraphs. Opening a new chat means having to explain who you are all over again, adjust the tone of voice, and re-explain the workflow—because Claude has no prior context about you. Claude’s built-in memory only records a few scattered notes, which is no substitute for a predefined set of rules, a knowledge base, and pre-packaged workflows.

This article guides you through building a 4-part “brain” for Claude: a working manual, a knowledge base, memory, and pre-packaged workflows. After reading, you will:

• Clearly understand what the 4 files do, how they differ, and their official names in Claude.
• Be able to set them up in one of two ways: Directly on the claude.ai web interface (no installation required), or via a local folder on your computer.
• Get a ready-to-use set of Vietnamese template files, 1 sample skill, and 3 prompts to have Claude generate them for you.
• Know how to turn a repetitive task that takes half a day into a single command.

The Big Picture: 4 Files, 4 Parts of the Brain

Imagine you have a brilliant employee who completely forgets everything from yesterday every morning they arrive at work. For them to do their job, every morning you must hand them 4 things: the company handbook, the project dossier, a notebook tracking what the boss previously instructed, and a set of standard operating procedures. The 4 files below are precisely those 4 things.

Method A: 5 steps to create a Project on claude.ai

Step 1: Create a New Project

Open claude.ai/projects (or click Projects in the left sidebar). Click the + New Project button in the top right corner. Give it a name, for example, “My Work Assistant”[4].

What is a Project? A dedicated workspace for Claude: Every conversation within it shares a common set of instructions and a document repository.

Step 2: Paste AGENTS.md into the instructions

On the Project page, click Set project instructions. Paste the entire contents of the AGENTS.md file (template in Part 8) populated with your information. Click Save instructions. From now on, all chats in the Project will follow this playbook.

Make 2 edits before pasting (since the web version lacks the memory.md file and the skills/ directory for Claude to write to automatically): In section 4, remove the steps “Read memory.md” and “skills/ directory”. In section 5, replace it with a single line: “Whenever we instruct you with something meant for long-term use, remember it using the Memory feature”.

Step 3: Drop documents into the knowledge base

In the Knowledge section on the right side of the Project page, click the + sign to upload the context.md file along with lengthy documents: price lists, project profiles, reports, brand guidelines (PDF, Word, or plain text files are all acceptable). On the web, documents are stored in a single location rather than having a tai-lieu/ folder, so in context.md, use the file name (e.g., “bang-gia.md”) instead of a file path. When the knowledge base approaches its limit, paid plans will automatically enable smart search (RAG) to accommodate more[4].

Step 4: Check that Memory is enabled

Memory is enabled by default for Free, Pro, and Max plans. To be sure, go to Settings to Memory and check if the Generate memory from chats toggle is turned on. Each Project has its own memory space, which does not mix with outside chats[5]. When you want Claude to remember something, simply say: “Remember this for us: …”.

Step 5: Enable code execution to use Skills

Go to Settings to Capabilities, then turn on Code execution and file creation. Skills require this feature to run[7]. How to install skills can be found in section 6.

Option B: Setting up the “brain” folder on your local machine

The complete folder structure looks like this (the template files in Part 8 are already arranged exactly this way):

bo-nao-claude/
├── CLAUDE.md         -->  bridge: Claude reads this first, loads the 3 files below
├── AGENTS.md         -->  employee handbook
├── context.md        -->  knowledge base table of contents
├── memory.md         -->  Claude self-logs lessons learned
├── tai-lieu/         -->  heavy documents: price lists, profiles, reports...
└── skills/
    └── viet-email-cong-viec/
        └── SKILL.md  -->  a packaged workflow

Why is CLAUDE.md necessary? Claude Code always automatically reads a file named CLAUDE.md when starting work. Inside this file are 3 lines: @AGENTS.md, @context.md, and @memory.md: The @ symbol means “load this file immediately”. Thanks to this, all 3 files are read automatically without requiring any reminders[2].

The Easiest Way to Install and Open the Folder

1. Download the Claude desktop app at claude.com/download (available for Mac and Windows), install it like any other application, and log in to your Pro account.
2. Create a folder named bo-nao-claude in an easily accessible location, such as in Documents. Create the files and subfolders exactly as shown in the diagram above, then copy and paste the contents of each template file from Part 8 into them.
3. Open each file—AGENTS.md, context.md—using Notepad (Windows) or TextEdit (Mac), and replace the bracketed placeholders [...] with your actual information. Save them with the proper .md extension (Notepad: Choose Save as type: All files; TextEdit: Go to Format, select Make Plain Text before saving).
4. In the Claude app, select folder mode and choose the bo-nao-claude folder (click the add folder button or drag and drop the folder in). Send your first prompt: “Read CLAUDE.md and summarize what you understand about us.”
5. If Claude summarizes it accurately, you’re all set. From now on, whenever you start working, simply open this exact folder.
6. If Claude summarizes it incorrectly or misses details, it hasn’t automatically read the file. In that case, send the above prompt at the beginning of each session before assigning tasks.

Using Claude Code (command-line version for developers): Open Terminal, run cd to the bo-nao-claude directory, then run claude. To let Claude Code automatically recognize skills, copy the skill directory into .claude/skills/ inside your project folder[6].

File 1: AGENTS.md, the employee handbook

Create this file first. It answers the question: “What is it like to work with this boss?” Before every assigned task, Claude reads this file to know how to do it.

What to include in AGENTS.md

• Who you are: Profession, company, clients, goals. Claude needs to know this to provide relevant, high-level advice.
• Working style: Short or long responses, whether to ask clarifying questions before proceeding, formatting preferences (tables vs. paragraphs).
• Tone of voice: What you like, what you dislike, accompanied by concrete examples.
• General workflow: What to read first, what to verify after.
• Memory rules: Instruct Claude to record lessons learned into memory.md (this is key to letting the third file grow autonomously).
• Strict constraints: Do not send emails automatically, do not delete files, do not fabricate data.

Be specific, don’t be vague

Claude reads these files as reference context; it is not forced to follow every single word verbatim. Anthropic’s documentation explicitly states: The more specific and concise the instructions, the more consistently Claude will follow them[2].

File 2: context.md, the raw material repository

AGENTS.md teaches Claude how to do it, while context.md gives Claude the raw materials to do it. Without these materials, Claude will only provide generic, textbook advice. With them, it provides advice tailored specifically to your product, your customers, and your project.

What to put here

• Products, services: Features, pricing, differentiators, competitors.
• Customers: Personas, pain points, frequently asked questions.
• Ongoing projects: Objectives, deadlines, who’s in charge.
• In-depth documents: Reports, internal policies, brand guidelines, contract templates.
• Internal jargon: Terms used only within your company.

Organization Tip: Use context.md as a Table of Contents

Don’t paste hundreds of pages into a single file. Keep long documents in a tai-lieu/ directory, while context.md contains only summaries and paths, like “Full price list: See tai-lieu/bang-gia.md”. Claude will read the table of contents first and only open that specific file when details are needed. This is both fast and saves usage limits.

File 3: memory.md, an assistant that can “read the room”

For example, if you say “don’t write emails in a playful, wordy tone,” Claude immediately jots down a line in memory.md. Next time, even in a new conversation, it will still remember. The longer you use it, the richer the file becomes, and the less you have to repeat yourself to Claude.

How Memory works

• On claude.ai (Approach A): The Memory feature automatically remembers details from conversations. You can view and edit them in Settings to Memory, under the Topics section. To save something immediately, say “Remember this for us”; to delete something, say “Forget about …”[5].
• In folders (Approach B): The “Memory rules” section in AGENTS.md instructs Claude to automatically write to memory.md. You can open the file to see everything and edit it manually with ease.
• In Claude Code: Built-in auto memory is available. Claude automatically saves 4 types of notes (about you, feedback, project, where to find information) into the MEMORY.md file, auto-loading the first 200 lines each session. Type /memory to toggle it on or off[2]. Using Claude Code with Approach B gives you both: MEMORY.md automatically managed by Claude, and memory.md under your own control, ready for you to open and read.

What a good memory line looks like

A good line includes a date, clearly states what it applies to, and serves as a reusable rule.

File 4: Skills, Teach Once and Use Forever

A Skill is an add-on instruction package for Claude to excel at a specific task, much like installing an app on your phone. You guide Claude through a workflow just once, then have it package that process into a skill to call upon whenever needed.

Why skills don’t “overload” Claude

Initially, Claude only reads the name and short description of each skill, which is about 100 tokens per skill (a token is a unit measuring the amount of text an AI processes). Only when the task you ask matches the description does it open the full instructions to read[6]. Therefore, installing 20 skills won’t slow Claude down. In return, the description must clearly state “when to use it”; if it’s written vaguely, Claude won’t know when to call it.

Create skills by chatting (no coding required)

Step 1: Do it once with Claude

Choose a task you do weekly, such as a weekly report. Work on it with Claude in a single chat, refining it until the result matches your expectations. Attach sample files and templates if available.

Step 2: Ask Claude to package it

Paste Prompt 2: Package the process into a skill (part 8). Claude will ask a few follow-up questions, then use the built-in skill-creator tool to write the SKILL.md file and package it into a .zip file[8].

Step 3: Install and test

Install following the instructions box below. Then open a new chat and ask for that specific task. Look at Claude’s thinking section—if you see the line “Using [skill name]”, the skill has been called[8]. If you are not satisfied with the result, simply message “Update skill: …”.

Easiest way to install a skill

1. Get the .zip file Claude just packaged for you. If you want to install the sample skill in Part 8: Copy the content of SKILL.md, paste it into the chat, and message “Please package this skill into a .zip file for us”.
2. Drag and drop the file into the Claude chat box, then message: “Please install this skill for us”. Claude will guide you or handle the rest.
3. Manual installation: Go to Customize > Skills, click the + icon, select + Create skill, choose Upload a skill, and select the .zip file[7].
4. Toggle on the skill in the list. If you do not see the Skills section yet, double-check in Settings > Capabilities whether “Code execution and file creation” is enabled.

Method B (folder): Create the folder skills/viet-email-cong-viec/ and save the content of SKILL.md there. The CLAUDE.md file has already instructed Claude to look for the skill in that location.

Putting It Together: From an Entire Afternoon Down to a Single Command

Take preparing a weekly report for a marketer as an example. It used to take an entire afternoon: gathering metrics, writing commentary, refining tone of voice, building tables, and drafting the email to the boss. Here is how the 4 files work together:

1. You type a single sentence: “Generate the weekly report from the attached metrics file and draft an email to the boss.”
2. AGENTS.md: Claude knows who you are, how your boss prefers presentations (conclusion first), and never to make up numbers.
3. context.md: Claude knows the quarterly goals, currently running projects, and compares metrics against targets.
4. memory.md: Claude remembers that last time you instructed: “don’t use pie charts,” “highlight 3 action items for next week.”
5. Skills: The “bao-cao-tuan” skill (which you packaged yourself following part 6) runs the exact steps you approved, while the “viet-email-cong-viec” skill drafts the email.

You receive a nearly finished report, only needing a quick review and a few tweaks. How much time you save depends on the task. The general rule of thumb: Any task you do a second time should be packaged into a skill.

Sample File Set and 3 Ready-to-Use Prompts

All pre-written in Vietnamese. Click Copy on each box, paste into a file or chat box, then replace the [...] placeholders with your information to use immediately.

The 4-File Framework and Bridge File

Method B uses all 5 files, arranged exactly as shown in the diagram in Part 2. Method A only requires AGENTS.md (pasted into Project instructions) and context.md (uploaded to Project knowledge).

# CLAUDE.md: Bridge file, Claude reads this file first in every session
# The three lines below starting with an '@' symbol tell Claude to load the other three files as well.

@AGENTS.md
@context.md
@memory.md

## About the skills/ directory

Each subdirectory in skills/ is a packaged workflow (a skill).
When the task we request matches the "description" of any skill, read that skill's SKILL.md file first and follow the exact steps inside.

# AGENTS.md: Working Handbook for [Your Name]

> This is your "employee handbook" (Claude). Read carefully before each task.
> Keep this file under 150 lines. Each rule should be 1 short, specific line.

## 1. Who I am

- Name: [John Doe], working as [Marketing Manager] at [Company X].
- Main customers: [Online shop owners, 25 to 40 years old, busy, not very tech-savvy].
- This year's goal: [Increase online channel revenue by 30%].
- Tasks we ask you for most often: [Writing emails, planning content, preparing weekly reports].

## 2. How you work with us

- Always reply in Vietnamese, address yourself as "we", and refer to us as "[anh/chị/bạn]".
- Get straight to the point. Conclusion first, explanation later.
- If critical information is missing, ask a maximum of 3 questions before proceeding.
- Do not fabricate data. If unsure, clearly state "Needs verification".
- If an answer is longer than 1 screen, divide it into sections with headings.

## 3. Tone of voice

- Write like a real person talking: Short sentences, simple words.
- Business emails: Polite, concise, maximum 150 words, ending with 1 clear proposal.
- Social media posts: Friendly, intriguing opening line, maximum 3 emojis.
- Avoid: Cliches ("In the 4.0 era..."), unnecessary English words, overly long rambling sentences.

## 4. Task handling process

1. Read context.md to understand the work context.
2. Read memory.md to recall what we have previously instructed.
3. If this task matches a skill in the skills/ directory, follow that skill precisely.
4. Once completed, self-check against the initial requirements before sending.
5. At the end of the answer, suggest 1 next step if applicable.

## 5. Memory rules (very important)

- Whenever we correct, critique, praise, or give an instruction that can be applied long-term, add 1 line to memory.md using the template: "- [date] Lesson learned".
- Before recording, re-read memory.md. If a similar idea exists, update the old line; do not create duplicates.
- Briefly notify us: "Remembered: ...".
- Do not write passwords, account numbers, or sensitive information into any file.

## 6. Things strictly forbidden

- Do not send emails, publish posts, or delete files without our consent.
- Do not modify original files in the tai-lieu/ folder. If edits are needed, create a new copy.

# context.md: Context Repository

> This file serves as the "table of contents" for the documentation repository. Keep it short (under 100 lines).
> Long documents should be placed in the tai-lieu/ folder and referenced only by their paths here.

## Work and Business

- Core products: [...]
- Pricing: [...]
- Key differentiators from competitors: [...]
- Main competitors: [...]

## Customers

- Persona: [...]
- Biggest pain points: [...]
- Frequently asked questions: [...]

## Ongoing Projects

| Project | Objective | Deadline | Details File |
|---|---|---|---|
| [Product Y Launch] | [500 orders in the first month] | [11/30] | tai-lieu/du-an-y.md |

## Reference Documents (read only when needed)

- tai-lieu/bang-gia.md: Full price list.
- tai-lieu/brand-guideline.md: Colors, logo, brand tone of voice.
- tai-lieu/bao-cao-q3.md: Q3 business metrics.

## Internal Terminology

- "Hot order": An order placed by a customer within 24 hours of seeing an ad.

## Email Signature

[Your name]
[Job title] | [Company]
[Phone number]