## Intro - The second installment of Vibe Coding terminology! In this second post, let's learn about what HOOKS, SubAgent (+ Claude Teams), and Harness are.🔥 > [!note] > - CLAUDE.md (=AGENTS.md) - Slash Commands - SKILLS - HOOKS ✅ - Subagent + Agent Teams ✅ - Harness ✅ - This time, too, the conceptual explanation will be based on Claude Code.🧑🏻‍💻
# 1. HOOKS 🪝 - The concept of a Hook has been around for a long time, but Claude Code was the pioneer in setting up and using hooks in an Agent. In the official documentation, Claude describes hooks as follows. > "are user-defined shell commands, HTTP endpoints, or LLM prompts that automatically run at specific points in the Claude Code lifecycle." - A HOOK, like a fishhook or a grappling hook, means "to catch something in its flow."🎣 What does it catch and where? You can think of it as **"catching the moment before or after a specific Event is executed,"**⚡️ such as when code, a session, or an AI task is finished. - For example, an Agent can do things like 'send a Slack notification when a task is complete,' or 'after creating or modifying code, automatically run a lint check and fix it until there are no errors.' - The biggest advantage of a HOOK is probably that, > Unlike Skills, which are probabilistic and may or may not run,
**it enables a deterministic setup where execution can be enforced every time without exception.** 👍 - As the originator, Claude Code currently supports hooks for a total of 18 events as of March '26 (and this number seems likely to grow), while **Antigravity does not have this feature**.😭 ~~Get to work, Google🦹‍♂️!~~ - It's said that tools like Gemini CLI, Cursor, OpenCode, and Codex CLI have started to support or are currently supporting HOOKS. ## Types of HOOKS - As of March 6, '26, Claude Code's [official docs](https://code.claude.com/docs/en/hooks-guide) support setting up hooks for a total of 18 events, including the ones below.👀 I've listed the 5 most commonly used events. | Event | When it fires | |---|---| | `PreToolUse` | Before a tool call is executed. Can be blocking. | | `PostToolUse` | After a tool call succeeds. | | `PostToolUseFailure` | After a tool call fails. | | `Notification` | When Claude Code sends a notification. | | `Stop` | When Claude completes a response. | ## How to Create a Hook with an example 📍 - A Hook can be registered by creating a `settings.json / settings.local.json` file under the .claude folder. The files used by the hook in settings.json are configured by registering the necessary files for each hook under the `.claude/hooks` folder. ![스크린샷 2026-03-07 오후 3.22.32](https://pub-08bdc1f395e14377951ac0c9c22ba795.r2.dev/posts/user_38T3Ty5uKeRndW9PUwTFHGYZxwP/g7tk8ms6qrj------------2026-03-07------3.22.32.webp) - As a simple example, a Hook to send a Slack notification to me when a task is completed can be set up as follows. 💬 `.claude/settings.json` ```json { "hooks": { "Stop": [ { "hooks": [ { "type": "command", "command": "/Users/peter/Desktop/claude_project/.claude/hooks/slack-notify.sh 'The task is complete.🧞‍♂️'", "timeout": 5 } ] } ] } } ``` - The above example is explained as follows. > [!tip] > - Event - `STOP`: Uses the end of a task as a trigger. > - The type is `command` > - The available types depend on the event type. [See Docs](https://code.claude.com/docs/en/hooks#prompt-based-hooks) > - The command content is to execute .claude/hooks/notify-done.sh as a bash script. > - Therefore, the `notify-done.sh` file must actually exist in the `.claude/hooks/` folder to run. > - `timeout` - Sets a 5-second timeout to avoid waiting indefinitely if Slack doesn't respond. - The notify-done.sh file contains code to send the "The task is complete.🧞‍♂️" message to Slack and is executed by the hook. - In conclusion, the file structure looks like this.
![스크린샷 2026-03-07 오후 3.31.42](https://pub-08bdc1f395e14377951ac0c9c22ba795.r2.dev/posts/user_38T3Ty5uKeRndW9PUwTFHGYZxwP/molwcpya0xj------------2026-03-07------3.31.42.webp) ## matcher 🏐 - Actually, the above example has a problem. The STOP event fires and triggers the hook even when the conversation ends without any code modification. To prevent unnecessary notifications and only send a Slack message when the Agent has actually modified/created code, I added a `PostToolUse` hook as shown below. - A concept not mentioned in the previous example is the [matcher](https://code.claude.com/docs/en/hooks#matcher-patterns). A Matcher acts as a filter that provides an additional layer of discernment (=matching) to trigger a hook only when a specific situation applies. ```json { "hooks": { "Stop": [ { "hooks": [ { "type": "command", "command": "/Users/peter/Desktop/claude_project/.claude/hooks/slack-notify.sh 'The task is complete.🧞‍♂️'", "timeout": 5 } ] } ], "PostToolUse": [ { "matcher": "Write|Edit", # 👈 matcher! "hooks": [ { "type": "command", "command": "/Users/peter/Desktop/claude_project/.claude/hooks/mark-changed.sh", "timeout": 3 } ] } ] } } ``` - The options you can write for this Matcher vary depending on the Event you are using. To see which matchers can be used for each event, [refer here](https://code.claude.com/docs/en/hooks#matcher-patterns).📍 - The hook set on the PostToolUse event above will only execute the mark-changed.sh script when code is created or modified, i.e., when it's a `Write|Edit` situation. This script's role is to plant a flag indicating a modification has occurred, which in turn signals to send a Slack notification. If the flag is not present, it means no code was modified or created, and no Slack message will be sent. - In the case of the Stop event hook written initially, there is no matcher at all. Like this, if the matcher is 1. completely empty, 2. `matcher: "*"` 3. `matcher: ""`
it is configured to **always execute in all Stop** situations. - It works very well. 😎✨ ![스크린샷 2026-03-07 오후 4.05.37](https://pub-08bdc1f395e14377951ac0c9c22ba795.r2.dev/posts/user_38T3Ty5uKeRndW9PUwTFHGYZxwP/zip7xxhwzcq------------2026-03-07------4.05.37.webp)
## 2. Subagents and Agent Teams 🦸‍♂️ - For Subagents and Agent teams, we will focus on the concepts. ### Creating a Subagent 🤖 - You can think of a Subagent as being used when a task is too large for the main agent to handle alone, or when multiple tasks need to be processed simultaneously. In other words, it is useful when you want to process work in parallel to quickly gather and report results. - So, while the main agent usually judges and divides tasks on its own, for faster results, you can explicitly write "use 10 parallel subagents to research independently as fast as possible," like Code Factory does. This makes the Subagents divide the work more quickly, and once they deliver their results to the main agent, it consolidates them to finish the job. ![스크린샷 2026-03-07 오후 9.08.36](https://pub-08bdc1f395e14377951ac0c9c22ba795.r2.dev/posts/user_38T3Ty5uKeRndW9PUwTFHGYZxwP/o4bl320jiy------------2026-03-07------9.08.36.webp) - As you can see below, multiple tasks are created. You can think of each task as a Subagent. - You can feel the speed increase linearly with the number of subagents used in parallel. Oh, and of course, you can also feel your tokens melting away at the same rate.😉
![스크린샷 2026-03-07 오후 9.51.47](https://pub-08bdc1f395e14377951ac0c9c22ba795.r2.dev/posts/user_38T3Ty5uKeRndW9PUwTFHGYZxwP/ybgnmluhhk------------2026-03-07------9.51.47.webp) ### Creating a Custom Subagent 🪄 - You can also create a separate Subagent suited for your own purposes! All you have to do is create a .md file with the agent's name under the `.claude/subagents` folder as shown below! After creating it, restart Claude Code to use it. You can also specifically address it in conversation or give commands with @Subagent-name. ![스크린샷 2026-03-07 오후 9.31.30](https://pub-08bdc1f395e14377951ac0c9c22ba795.r2.dev/posts/user_38T3Ty5uKeRndW9PUwTFHGYZxwP/2tjnnki9kv2------------2026-03-07------9.31.30.webp) - In fact, the format is very similar to creating a Skill, as we saw in the previous post. You write the Metadata (=Frontmatter) in YAML format with predefined keywords, and the Body in markdown. - For a Subagent, only `name` and `description` are required, but you can configure many more things, such as which model to use, what tools/skills/MCPs it can use, etc. For more details, [refer here](https://code.claude.com/docs/en/sub-agents#supported-frontmatter-fields). ✔ - In Antigravity, as of March 2026, the only available Subagent is the Browser Agent, which can launch Chrome to check its operations during development. In OpenAI Codex, it's reportedly available under the name Multi-agent. ### Agent Teams 🤝 - The Agent Teams feature is also currently only available in Claude Code. Other services, including Codex, do not have a feature like Agent Teams at present.🫢 - The biggest difference between Subagents and Claude Teams is as follows. ![subagents-vs-agent-teams-dark](https://pub-08bdc1f395e14377951ac0c9c22ba795.r2.dev/posts/user_38T3Ty5uKeRndW9PUwTFHGYZxwP/2onyboyl3t5-subagents-vs-agent-teams-dark.webp) - A subagent simply completes its assigned task and returns the result to the main agent, but the biggest difference is that an Agent Team shares a task list, and each teammate can talk to and exchange opinions with one another. - To give a more intuitive example, if you were creating a bulletin board feature, you would need at least three agents for Front-end, Back-end, and DB, plus a fourth agent for Test & Lint validation. - In this situation, **each agent must check the status of the others to ensure the features they create work well together, waiting for their turn to build.** This would be especially true in a monorepo, and I think this Claude Team feature would be even more useful. ### How to Activate Claude Team📍 - Activating Claude Team is simple. Just add the following to the `.claude/settings.json` file in the scope of the environment you want to use, and you're done! ```json ... "env": { "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1" }, "teammatemode": "tmux" } ``` - The `teammatemode` option has two choices, "tmux" and "in-process". This setting determines whether you view the teammates running separately in the Claude Code terminal as tabs (in-process) or in a split-window tmux style. - Once activated, if you restart Claude Code, for example, just giving the command **'Create and run a team of 4 members'** is enough to see it form a team and start working. For those curious about what it looks like in practice, please refer to Code Factory's [video](https://www.youtube.com/watch?v=Gb2VMWrUmZ0&t=647s).😊 ~~It's not that I'm too lazy to write it...~~
## 3. Harness Engineering? - Phew, we've finally made it to the last keyword!🔥 The keyword `Harness` comes from horse tack. It's a collective term for equipment like the reins, saddle, bit, and yoke used to control a horse. - So, if you can't handle these wild AI Agents properly, they might just start developing in crazy directions, and you can feel that it wouldn't be strange if it turned into the movie Terminator. It seems like a fitting word for this age of singularity. - In that sense, `Harness Engineering` means **designing systems to control powerful yet unruly AI in the right direction**. - How do you design them? It wouldn't be a stretch to interpret the meaning of Harness Engineering as of March 2026 as meticulously designing things like Commands, Skills, and Hooks, which we've looked at together over two posts, to create Plugins. ### Recommended Harness Plugins 🔖 - Here are some Harnesses I would like to recommend based on my personal judgment. #### 1. [Superpowers](https://github.com/obra/superpowers) 💪 - This is an incredible Harness with 73k stars on Github. - Personally, the most impressive part is that it's a Harness that makes the AI code based on TDD and finish the tests on its own.👀 So cool. I haven't seen any other TDD-based harnesses yet. - Available for Claude Code, Codex, Cursor, OpenCode. #### 2. Oh-my-family 🏠 - If you use Claude Code, I recommend [Oh-my-Claudecode](https://yeachan-heo.github.io/oh-my-claudecode-website/)🔥, if you use Codex, [Oh-my-codex](https://github.com/Yeachan-Heo/oh-my-codex)🤖, and if you use OpenCode, [Oh-my-openagent](https://github.com/code-yeongyu/oh-my-openagent)🪨. - These are all brilliant harnesses with well-made Subagents, Commands, and Skills.✨ #### 3. [Ouroboros](https://github.com/Q00/ouroboros?tab=readme-ov-file) 🧘 - It elicits the essence of what the user wants through a Socratic questioning method. It quantifies things like ambiguity or states where the agent gets stuck in a loop asking the same question (Oscillation), making the agent work very smartly, persistently, and properly. An amazing harness that I've fallen for...🫢 - For Claude Code only.
## Conclusion 🍳 - There are so many experts out there running wild(?). This has become an era where those with deep humanistic insight and contemplation, who can implement philosophical thoughts into code and apply them to AI, hold the greatest power. You can feel that these people are enjoying this Vibe Coding era like fish in water. ~~So jealous it makes me mad😭~~ - For an ordinary person like me, it's an era where I'm just busy trying to keep up with what they create, but I think this level of basic knowledge is essential to quickly understand how those plugins and skills, in other words, Harness Engineering, are structured and configured. - Now that we know the basics, we'll be able to build our own Harnesses by weaving together Commands, Skills, Hooks, etc., tailored for our own purposes or for our company's team.😎

✨HAPPY BUILDING!👷‍♂️ ### References - [datacamp.com](https://www.datacamp.com/tutorial/claude-code-hooks) - [Claude Code Docs - Hooks](https://code.claude.com/docs/ko/hooks) - [Code Factory](https://www.youtube.com/@codefactory_official)

Vibe Coding Terms - Part 2

0 comments