People ask us which Claude Code features are worth the time to learn. These are the twelve we use on client work, in three groups. The three we set up first are at the end.
Doing more work
Subagents
Separate workers that each take one part of a job at the same time. In a codebase review, ten of them can each read a different file. Each one has its own context, so the main session stays short.
Workflows
A script that sends work out to many subagents and collects the results. The order of the steps is fixed in code, which makes a weekly job run the same way each week.
Scheduled tasks
Agents that run on a schedule in the cloud: a weekly digest, an end-of-day report, a data cleanup overnight. The work starts on time whether or not anyone is logged in.
Loop
One prompt that repeats on an interval until someone stops it or a condition is met. It suits a long job that needs watching, or a system that cannot send a notification. For a few hours of use it is quicker to set up than a scheduled task.
Checking the work
Plan mode
A read-only mode. Claude studies the code and writes a plan, and the files stay as they are until a person approves it. A wrong idea gets fixed in the plan.
Memory and project files
Standing instructions and conventions that carry over between sessions. A team writes down how it works once, and every session starts with that.
Checkpoints
A saved state that a session can rewind to when a change goes wrong. Claude Code saves them as it works.
Verification
A second agent checks the first agent’s work before a person sees it. It adds tokens and time, and it suits any change that would be costly to get wrong.
Reaching other tools
Remote sessions
A session that starts from a phone, an email, or a webhook. It runs in the cloud, and the result arrives when it finishes.
MCP connectors
The Model Context Protocol (MCP) connects Claude to other systems, such as a CRM, shared files, or an internal API. Claude can read and update the data in them. This is how Claude does business work as well as code work.
Skills
A skill packages the instructions and conventions for one job, such as a code review or a release note. Once a skill is in a project, every session can use it.
Hooks
Scripts that run at fixed points: before a tool runs, after it runs, when a session ends. They run every time, whatever Claude decides, which makes them a good place for guardrails and audit logs.
How they fit together
The features are built to combine. One setup uses six of them. A scheduled task starts a workflow. The workflow sends subagents out to business systems through MCP connectors. A second agent checks the results, and a hook writes the audit log.
The three we set up first
We set these up in the same order on almost every project. Memory and project files come first, and every session then starts with the same context. Plan mode is second: a person reads the plan before anything changes. Verification is third, so the work is checked before anyone relies on it.
Subagents and scheduled tasks come after those three. They raise the amount of output, and by then the team has a way to check it.