Claude Code Has 50+ Features Most People Never Discover

Claude Code Has 50+ Features Most People Never Discover

I use it eight or more hours a day. These are the things I wish I had known on day one.

The One Thing That Changed Everything

Start every task in plan mode. Shift+Tab twice.

Let it read the codebase first. Let it break your goal into steps. Review the plan. Then switch to auto-accept and let it execute end to end.

This one workflow cut my revision cycles by roughly 80%.

The reason it works is simple arithmetic on the cost of being wrong. A bad assumption caught in a plan costs one sentence to correct. The same assumption caught after execution costs a full revision cycle, plus the work of unpicking a diff that went in the wrong direction.

Most frustration with agentic coding tools comes from skipping the cheap correction and paying for the expensive one.

The Three Configuration Layers

Most people use the tool with none of these configured, then wonder why it does not know their conventions.

CLAUDE.md gives it context. A project file loaded into every session. Conventions, architecture decisions, the commands your project actually uses. Anything you would otherwise re-explain every day belongs here.

Skills give it capability. Packaged instructions for a recurring task, written once and invoked by name. The difference between a skill and a good prompt is that a skill lives in the repository, gets versioned with the code, and works the same way for everyone on the team.

Hooks give it enforcement. Commands the harness runs automatically at defined points. Formatting, linting, tests. Because the harness executes them rather than the model deciding to, they happen reliably.

That last distinction matters. Anything you need to happen every time belongs in a hook, not in an instruction. Instructions are followed most of the time. Hooks run.

Model Selection

Haiku is fast and cheap, and it is the right choice for mechanical or high-volume work where the task is well-specified.

Sonnet is the balanced default for most day-to-day work.

Opus is for the hardest reasoning: architecture decisions, subtle debugging, anything where being wrong is expensive.

The common mistake is using the most capable model for everything, which is slower and more expensive than it needs to be for tasks that are genuinely mechanical.

Where The Real Leverage Is

MCP servers connect it to the tools you already pay for. This is the feature that turns a coding assistant into something that operates your stack, because it means the agent can call your CRM, your data provider, and your sender directly.

Scheduled agents run monitoring jobs without you present. Useful for anything you would otherwise check manually on a cadence.

Sub-agents split a task so each part gets its own instructions and its own constraints. Particularly effective when the pieces have different rules, like generating headlines under one character limit and descriptions under another.

What It Actually Is

This is not a chatbot you paste code into.

It reads your codebase, writes files, runs tests, and completes whole features while you review. That is a different category of tool, and using it like a chat window is why some people conclude it is overrated.

The mental shift is treating it like a capable colleague who has just joined and does not yet know your conventions. You would not expect that person to produce good work without context about the project, and you would review their pull requests rather than watching them type.

A Realistic Caveat

It is confidently wrong sometimes. Review diffs the way you would review a colleague's work, especially on anything touching data, money, or production.

The failure mode is not obvious nonsense. It is plausible, well-structured code built on a misunderstanding of your system, which is precisely the kind of error that survives a quick glance.

Plan mode helps here too, because a misunderstanding shows up in the plan before it becomes a diff.

Configuration At A Glance

Frequently Asked Questions

What is plan mode and why does it matter so much?

Shift+Tab twice puts it into a mode where nothing is written to disk. It reads the codebase, proposes an approach, and waits. The value is arithmetic: a wrong assumption corrected in a plan costs one sentence, while the same assumption corrected after execution costs a full revision cycle plus unpicking the diff.

When should I use an instruction versus a hook?

Anything that must happen every time belongs in a hook. Instructions are followed most of the time; hooks are executed by the harness, so they run regardless. Formatting, linting, and tests are hook material.

How do I choose between Haiku, Sonnet and Opus?

Haiku for mechanical, well-specified, high-volume work. Sonnet as the balanced default. Opus for the hardest reasoning, architecture decisions, and subtle debugging. Using the most capable model for everything is slower and more expensive than the work requires.

What is the difference between a skill and a well-written prompt?

A skill lives in the repository, is versioned with the code, and works identically for everyone on the team. A prompt lives in one person's history and disappears when they do.

What is the most common mistake new users make?

Treating it like a chat window. It reads your codebase, writes files, runs tests, and completes whole features. Pasting a snippet in and asking for advice uses a small fraction of what it does, and it is the main reason some people conclude the category is overrated.

We build AI-native pipeline systems and you pay per qualified meeting, not a retainer. No paying for activity. You only pay when we book you a qualified sales meeting. See if you qualify.

Feature descriptions reflect Claude Code as of 2026. Check the official Anthropic documentation for current capabilities and model availability.