## Intro - CLAUDE, Antigravity, Codex, OpenCode, etc. Even if the AI Coding Agents you use are all different, once you step into this world, there are certain words you're bound to hear—words that make you feel like you'll fall behind if you don't know them. ~~(There are so many...🙃)~~ > [!note] > - CLAUDE.md(=AGENTS.md) ✅ > - Slash Commands ✅ > - SKILLS ✅ > - HOOKS > - Subagent > - Harness - Today, let's look at the first three of these. The conceptual explanations will be based on Claude Code, and we'll proceed in the order of Antigravity's setup method. Antigravity's setup is almost identical or very similar to that of other AI Coding Agents like OpenCode and Cursor, so please take note ✔
## 0. Token 🎟️ - As you may already know, almost all .md files are written in English. This is to use fewer tokens (= to save money). - Using English results in 6 tokens.
![스크린샷 2026-02-27 오전 7.59.06|400](https://pub-08bdc1f395e14377951ac0c9c22ba795.r2.dev/posts/user_38T3Ty5uKeRndW9PUwTFHGYZxwP/a74ztl3sv3g------------2026-02-27-------7.59.06.webp) - The same phrase in Korean is 12 tokens!
![스크린샷 2026-02-27 오전 7.58.42|400](https://pub-08bdc1f395e14377951ac0c9c22ba795.r2.dev/posts/user_38T3Ty5uKeRndW9PUwTFHGYZxwP/yy604uywz1------------2026-02-27-------7.58.42.webp) - If you compare sentences with the same meaning in English and Korean on OpenAI's [Tokenizer], you can see that Korean uses at least 1.5 to 2 times more tokens than English. A Token is the 'smallest unit for an LLM to process text by converting it into numbers'. - Typically, when an LLM is trained, it understands text by grouping frequently co-occurring character patterns in the training data into single tokens. Since the amount of training data for English is overwhelmingly large, English can interpret relatively large semantic units as a single token. In contrast, for foreign languages like Korean, the training data is relatively small, which leads to this phenomenon. - So, keep in mind that writing .md files for Commands, Skills, Hooks, etc., in English is a way to reduce the tokens used each time they are executed! ✔️
## 1. CLAUDE.md / AGENTS.md - CLAUDE.md / AGENTS.md is a `globally loaded context that is always present`. Therefore, **it's best to write down common rules for projects/teams to ensure they are always recognized without exception, or to focus on writing only the most crucial information**. Otherwise, sentences irrelevant to the actual command will just incur token costs, confuse the AI, and degrade its performance. - Antigravity uses GEMINI.md, but you can also use AGENTS.md. - Usually, CLAUDE.md/AGENTS.md contains information about the project, the specifications it uses, code style, commands, architecture, and important points. Simply put, you can understand it as writing: - `What` - What you are building and with what stack - `Why` - Why other files and the file structure exist within this project (Structure) - `How` - How tasks must be performed. - Here is an example for a Next.js project. ```markdown # Project: ShopFront Next.js 14 e-commerce application with App Router, Stripe payments, and Prisma ORM. ## Code Style - TypeScript strict mode, no `any` types - Use named exports, not default exports - CSS: Tailwind utility classes, no custom CSS files ## Commands - `npm run dev`: Start development server (port 3000) - `npm run test`: Run Jest tests - `npm run test:e2e`: Run Playwright end-to-end tests - `npm run lint`: ESLint check - `npm run db:migrate`: Run Prisma migrations ## Architecture - `/app`: Next.js App Router pages and layouts - `/components/ui`: Reusable UI components - `/lib`: Utilities and shared logic - `/prisma`: Database schema and migrations - `/app/api`: API routes ## Important Notes - NEVER commit .env files - The Stripe webhook handler in /app/api/webhooks/stripe must validate signatures - Product images are stored in Cloudinary, not locally - See @docs/authentication.md for auth flow details ```
### File Creation Location 🗂️ - It is divided as follows, depending on whether it is global or project-specific. ~/.claude/CLAUDE.md                        # 👈 Global
project-root/.claude/CLAUDE.md     # 👈 Project-specific ~/.agents/AGENTS.md                        # 👈 Global
project-root/.agents/AGENTS.md     # 👈 Project-specific ### Where CLAUDE.md is located (3 levels)📍 - CLAUDE.md operates on three levels, which is an important aspect that most tutorials don't cover properly. `Level 1: Global` — ~/.claude/CLAUDE.md - This applies to all projects. It contains personal settings like your preferred coding style, language settings, and frequently used tools. ✍️ `Level 2: Project` — project/CLAUDE.md - This applies only to the specific repository. It contains the project architecture, coding conventions, build commands, etc. This file is shared with the team via Git. 🔄 `Level 3: Folder` — project/src/CLAUDE.md - This applies only within a specific directory. It's not loaded entirely at the start of a session but is loaded on-demand when Claude reads a file in that directory. ✅ `Priority`: - **Folder > Project > Global**. In other words, you can understand it like this: ``` ~/.claude/CLAUDE.md → "Use JavaScript" project/CLAUDE.md → "Use TypeScript" ← This one wins in this project project/src/CLAUDE.md → "Turn on strict mode" ← This one wins in this folder ``` ![claudemd-example](https://pub-08bdc1f395e14377951ac0c9c22ba795.r2.dev/posts/user_38T3Ty5uKeRndW9PUwTFHGYZxwP/f8gbiqcjijg-claudemd-example.webp) - For example, if you set 'Use Python' at the Global level and 'Use JavaScript' at the Project level, it will use JavaScript. - This concept is the same for any tool that uses AGENTS.md, such as Amp, Codex, Cursor, Devin, Factory, Gemini CLI, etc... However, Antigravity only applies the first two levels of hierarchy; the third level, the file-based hierarchy, is said to be meaningless. ~~(Get to work, Google)~~
### CLAUDE.md Best Practices ✨ - There are several guides and papers on how to best use CLAUDE.md or AGENTS.md. The content below is what is generally considered accepted as of March '26. #### `1. Less (instructions) is more.` 😌 - It is recommended to write it within 500 lines if possible (under 300 characters is best). Instructions over 500 lines can actually degrade performance. - Once it starts exceeding 500 characters, a 'Lost in the Middle' phenomenon occurs, where it only remembers the beginning and end, forgetting the middle part. #### `2. Progressive Disclosure` 🔄 - Don't write all the information in CLAUDE.md. Instead, design it so that the model can refer to it and load it into memory as needed. - For example, you can create a separate directory as shown below, write the instructions, and then in CLAUDE.md, only mention which file to look at in which situation. This allows the AI to decide and read it on its own when needed. ```bash agent_docs/ |- building_the_project.md |- running_tests.md |- code_conventions.md |- service_architecture.md |- database_schema.md |- service_communication_patterns.md ``` - In other words, it's a way of **progressively disclosing information to the AI**. #### `3. Agent is not a linter.` 🙅 - Linters and formatters are for catching TypeScript errors or syntactical errors that could be a problem during builds. Making the AI do this, especially by writing it in CLAUDE.md/AGENTS.md, means you're using a feature unnecessarily expensively. - When there are already many tools specialized for a specific function like this, it's better not to write it in CLAUDE.md but to automate it by setting up a Command, which we'll look at next, or by using Hooks, which we'll discuss in the next article.
## 2. Slash Commands💂‍♀️ - Slash Commands are also simply referred to as `Commands`. They are like functions that allow you to execute a pre-made feature with a `/your_command` command. #### Claude Code 1. If you want to use a command globally for any project, create a directory named `commands` under `~/.claude/`. To use it in a specific project, create a folder named `.claude` in the project's root path, and then create a `commands` folder inside it. mkdir -p ~/.claude/commands mkdir project-root/.claude/commands - Next, just create a `your_desired_skill_name.md` file under the Commands folder, and you're done! - I used the Obsidian MCP setup from my previous post to create a `/today` command as shown below.
![스크린샷 2026-03-01 오전 10.36.23|700](https://pub-08bdc1f395e14377951ac0c9c22ba795.r2.dev/posts/user_38T3Ty5uKeRndW9PUwTFHGYZxwP/8et9aqzo7i------------2026-03-01-------10.36.23.webp) - Hehe, it works well. The speed is not bad either.
![스크린샷 2026-03-01 오전 10.39.44|500](https://pub-08bdc1f395e14377951ac0c9c22ba795.r2.dev/posts/user_38T3Ty5uKeRndW9PUwTFHGYZxwP/18mf2vg42wp------------2026-03-01-------10.39.44.webp) #### Antigravity - In Antigravity, as shown in the picture below, you can set up Commands by clicking the three dots in the upper right, selecting `Customizations` from the menu, and then clicking on `Workflows`. - You can select the environment where the command will apply by clicking Global/Workspace and then create the command. ![스크린샷 2026-02-28 오후 5.58.48|750](https://pub-08bdc1f395e14377951ac0c9c22ba795.r2.dev/posts/user_38T3Ty5uKeRndW9PUwTFHGYZxwP/sr48a0oq4ip------------2026-02-28------5.58.48.webp) - Then, as you can see below, it automatically creates a `commands` directory under the `.agents` folder and generates an .md file with the command name you set. = In other words, you can just create the `agents/commands` folder and the .md file yourself!🤗
![스ᄏṛᅵᆫ샷 2026-02-28 오후 8.37.58|650](https://pub-08bdc1f395e14377951ac0c9c22ba795.r2.dev/posts/user_38T3Ty5uKeRndW9PUwTFHGYZxwP/b739bvd9wmt------------2026-02-28------8.37.58.webp)
## 2. SKILLS 🧰 - Skills are truly fascinating. As we saw earlier, CLAUDE.md or AGENTS.md are always executed and consume tokens in every conversation, but **once a Skill is set up, the Agent will execute it automatically when it deems it necessary for a given situation.** - This concept is called `Progressive Disclosure`. That's right, it's the same concept mentioned when explaining CLAUDE.md. Skills implement this concept by dividing it into three main levels, as shown in the image below. ![스크린샷 2026-03-03 오후 4.27.02|800](https://pub-08bdc1f395e14377951ac0c9c22ba795.r2.dev/posts/user_38T3Ty5uKeRndW9PUwTFHGYZxwP/bdw9g1e42cq------------2026-03-03------4.27.02.webp) `Level 1. Metadata` - In the Context. In other words, this is the part that is always reflected in Memory. The rest is not loaded into memory until it's needed based on the user's prompt, at which point the remaining sections are read. - The Frontmatter is written in YAML format, enclosed by `---` at the top and bottom, and follows the rules below. - Basically, `name` and `description` are required. The agent analyzes the user's intent, and if the description matches, it automatically loads the corresponding skill. > [!note] > `name`: > - Maximum 64 characters > - Must contain only lowercase letters, numbers, and hyphens > - Cannot contain XML tags > - Cannot contain reserved words: "anthropic", "claude"
> > `description`: > - Must be non-empty > - Maximum 1024 characters > - Cannot contain XML tags > - Should describe what the Skill does and when to use it > - For complete Skill structure details, see the Skills overview. - For more options, see [here](https://platform.claude.com/docs/en/agents-and-tools/agent-skills/overview#skill-structure). `Level 2. Instructions` - This is the section where specific instructions are written in Markdown. It is loaded once the decision to use the skill is made. `Level 3. Bundled files` - In Level 2, if the Markdown content contains a link like 'Refer to this file,' it will go one depth deeper to read that file, thus entering Level 3. Using the same concept, 3+ levels can also be implemented. ### SKILL Example ✔ - When creating a skill to process PDFs, you can create a folder named `pdf/` and generate files with the following structure underneath it. ``` pdf/ ├── SKILL.md # Main instructions (loaded when triggered) ├── FORMS.md # Form-filling guide (loaded as needed) ├── reference.md # API reference (loaded as needed) ├── examples.md # Usage examples (loaded as needed) └── scripts/ ├── analyze_form.py # Utility script (executed, not loaded) ├── fill_form.py # Form filling script └── validate.py # Validation script ``` - The SKILL.md looks like this, and it's an example of a Level 3 skill where it refers to additional documents like `/reference.md` and `/forms.md` in its body. 😎 ![agent-skills-bundling-content|auto](https://pub-08bdc1f395e14377951ac0c9c22ba795.r2.dev/posts/user_38T3Ty5uKeRndW9PUwTFHGYZxwP/nos0e53lbpl-agent-skills-bundling-content.webp)
### File Creation Location 🗂️ - So, where do you create this `pdf` folder? It's the same as creating `commands`, you just need to name the folder `skills`! Isn't that easy?👨🏻‍🎨 #### CLAUDE ~/.claude/skills/
    └── my-skill/
               └── SKILL.md
               └── And so on.. project-root/.claude/skills/
    └── my-skill/
               └── SKILL.md
               └── And so on.. #### Antigravity ~/.gemini/antigravity/skills/
    └── my-skill/
               └── SKILL.md
               └── And so on.. project-root/.agent/skills/
    └── my-skill/
               └── SKILL.md
               └── And so on..
## SKILL Store 🏪 - The era of showing off and sharing skills created this way has begun. I'll introduce three representative examples below, but so many developers worldwide are creating and sharing great skills, so it's a good idea to keep an eye on X or Threads.🧐 ### 1. Vercel Skills - A prime example is [skills.sh](https://skills.sh), created by Vercel. Just by visiting the homepage, you can see eye-catching skills you'll want to install and use right away, like a skill for finding skills, `react-best-practices`, and `design-guide`.😎 - `find-skills`, for example, is a skill that lets you find and use skills you don't have installed. Hehe, be sure to install it. It supports almost all various Agents like Antigravity, Claude Code, etc. ![스크린샷 2026-03-03 오후 7.22.44|auto](https://pub-08bdc1f395e14377951ac0c9c22ba795.r2.dev/posts/user_38T3Ty5uKeRndW9PUwTFHGYZxwP/84fy53wo8de------------2026-03-03------7.22.44.webp) ### 2. [Anthropic Skills](https://github.com/anthropics/skills?tab=readme-ov-file) - The official Anthropic skill set 👀. So it's exclusive to Claude Code. There are many great-looking skills like Creative & Design skills.🤤 ~~Just install it already!!~~ ![스크린샷 2026-03-10 오전 12.25.55](https://pub-08bdc1f395e14377951ac0c9c22ba795.r2.dev/posts/user_38T3Ty5uKeRndW9PUwTFHGYZxwP/hzhpqzwdbsc------------2026-03-10-------12.25.55.webp) ### 3. [Antigravity Awesome Skills](https://github.com/sickn33/antigravity-awesome-skills/tree/main) - It can run on various tools like Claude Code, Codex, Cursor, etc. They say there are 1234+ SKILLS🫢, so take a look! ![스크린샷 2026-03-10 오전 12.23.10](https://pub-08bdc1f395e14377951ac0c9c22ba795.r2.dev/posts/user_38T3Ty5uKeRndW9PUwTFHGYZxwP/tiuqkhjpmn------------2026-03-10-------12.23.10.webp)
## Conclusion 🍳 - It seemed daunting when I didn't know about it, but after organizing my thoughts, I realize these are all just tools born from the following ideas: > "I want to achieve maximum efficiency with fewer tokens for the same question."
> "I want to conveniently save and call frequently used commands."
> "I wish it would just take care of things intelligently, neatly, and sensibly on its own."
✨ HAPPY BUILDING! 👷
### References - [CLAUDE API Docs](https://platform.claude.com/docs/en/agents-and-tools/agent-skills/best-practices) - [Lost in the middle](https://arxiv.org/abs/2307.03172) - [Writing a good CLAUDE.md](https://www.humanlayer.dev/blog/writing-a-good-claude-md) - [Builder.io VISHWAS GOPINATH](https://www.builder.io/blog/claude-md-guide) - [Evaluating AGENTS.md: Are Repository-Level Context Files Helpful for Coding Agents?](https://arxiv.org/abs/2602.11988) - [Dheeraj Sharma](https://genaiunplugged.substack.com/p/claude-code-skills-commands-hooks-agents)

Vibe Coding Terms - Part 1

0 comments