Transcription
Coding agents like Cloud Code are fundamentally changing how software gets built. In this video, we cover the top 12 Cloud Code features that matter most.
Every time you start a new session, Cloud knows nothing about your project structure and your coding preferences. CloudMD fixes this. It is a markdown file containing things like coding preferences and project structure sitting in your project route. Cloud reads it at the start of every session using it as a memory. You write rules like always write unit tests and cloud follows them in every session. There is no required format for cloud.md. Just keep it short and human readable. If you type a /init in your cloud session, it automatically creates cloud.MD for your current project.
Coding agents are powerful but risky since they can edit files and run commands on your machine. By default, cloud code asks for your approval before editing files or running bash commands. Permissions let you customize this behavior. You can preapprove safe actions like running tests or committing code and block dangerous ones like deleting files. Inside that session, type /ermissions. This opens an interactive menu where you can add or remove tools from your allow list. Permissions are the balance between a speed and safety. Start conservative and gradually open up as you build trust with the specific workflows.
Coding right away is tempting, but on complex tasks, Cloud can edit the wrong files and waste tokens. Plan mode separates planning from doing. Press shift plus tap to switch between plan mode and normal mode. When enabled, Claude can read your files, ask questions, and propose a step-by-step plan. Only read only tools are allowed in this mode. So, Claude cannot make changes. Once you review and approve the plan, you switch back to normal mode and Claude executes it.
You asked Claude to refactor a module. After editing many files, you realize the approach is wrong. Now you are stuck manually undoing everything. Checkpoints solve this. Before each edit, Cloud automatically snapshots your files, so you can rewind to any earlier state if anything goes wrong. Type a /re in cloud session to see a list of checkpoints with timestamps and descriptions. Select the one you want to restore. Choose what to rewind and you are back to that exact state. Checkpoints let you try bold experiments and explore risky ideas without losing working code.
Some workflows need the same detailed instructions every time. Typing them out repeatedly is tedious. Skills are predefined instructions Cloud can load on demand. Each skill is a skill.md file with a name, description, and instructions. At session start, we provide cloud with a list of available skills and it auto invokes the right skill when needed. Skills turn repeated workflows into repeatable sharable automation. Build a skill once and your entire team benefits from it without memorizing long prompts.
Some actions should happen every single time, like formatting code after every code changes. Clot code works through a workflow loop, deciding when to call tools and repeating that process as needed. Hooks are scripts that run automatically at a specific points you choose in that workflow, such as before or after a tool run. Use hooks for deterministic actions that must always happen. Formatting, security checks, or logging.
Cloud code can read files and run bash commands. But what about external tools like your Figma files or Slack content. MCP is an open protocol that allows anyone to build and expose tools to agents. You add an MCP server and cloud gets access to all the tools it exposes. MCP gives you immediate access to thousands of publicly available MCP servers.
You have spent hours configuring the perfect setup. Custom skills, hooks, sub aents, MCP servers. Now your teammate wants the same setup. Do you walk them through every config file manually? Plugins are bundles that package skills, hooks, agents, MCP servers, and metadata into a single installable unit. You create a plug-in, publish it to a marketplace or a Git repository, and your teammates install it with one command.
Cloud Code works within a fixed context window of about 200,000 tokens. As your conversation grows, that window fills up. When it gets close to capacity, Cloud can compact the conversation, preserving key decisions while freeing up space. Run a /context to see exactly what is consuming your context window. CloudMt files, skills, MCP tool descriptions, and conversation story. If anything critical is close to being lost, run/compact with instructions before it happens automatically.
Some actions are so common that typing them every time slows you down. A slash commands act like keyboard shortcuts for cloud code. They give you a quick way to trigger common workflows such as checking costs, managing context, or clearing your current session. You invoke by typing a slash followed by the command name.
For complex tasks, doing everything in one session gets messy fast. A sub agent is a separate cloud session for one specific job. Claude splits a complex task into a smaller focused pieces and hands them off. The sub agents work independently, then return a short summary to the main session. Type a / agents and select create agent. Give it a name and a prompt focused on a specific area like security review. Sub aents keep your main conversation clean and your context window lean. Use them for any task that involves heavy exploration. parallel work or a specialized expertise.
Now, cloud code feels simple and structured. The best way to learn is to open a terminal, type cloud and start building. If you like these videos, you may like our AI engineering cohort as well. Everything is project based, so you are building from day one. Thousands of learners are already in. We cover fundamentals like reasoning LLMs and agents and you'll build real systems like deep research workflows, rags and web search agents with MCP. Check out the full curriculum link in the description.