I've spent hundreds of hours in Claude Code, and I have an uncomfortable truth for you: most "best features" rankings are useless, because they measure the wrong thing. What matters isn't which feature is the flashiest or which the most people use — it's which one actually changes your working day. Those are two completely different lists.
I'll show you the twelve features that make that difference for me, and tell you why. But first a caveat I want you to take seriously.
The ranking depends on how you work
There's no single, objective ranking of Claude Code features — and anyone who claims otherwise is selling you an illusion. What's my number-one feature may be a curiosity to a programmer building software, and something a programmer treats as essential may not touch my work at all.
The way I work is knowledge work and automation: organizing, researching, writing, running repetitive processes. I don't build large software products. So my ranking is tilted toward that kind of work — and it's good that you know it. Take it with a grain of salt and treat it as a map, not a verdict. The goal is just one thing: to show you a few features you may not know, or aren't using the way you could.
Before I get to the list, let me pin down two terms that will recur. Context is the model's short-term memory within one conversation — everything Claude "sees" at a given moment. It's finite, so it has to be managed. An agent is Claude working on its own according to the instructions you leave it, instead of answering each prompt separately. The rest of the jargon I'll explain as we go.
The top five — the features that change the most
01Skills
This is my number-one feature, and I have no doubt about it. A skill is simply a recipe. Just as you don't make lasagne from memory every time — you open the recipe and follow the steps — you give the agent a ready-made recipe for a task you do often.
A skill is a text file (markdown) that tells Claude exactly how you like to do a given thing: prepare a presentation, make a thumbnail, trim a video, hand a session on. So you don't have to write the same long prompt every time — Claude reads the recipe and does it consistently the same way. Skills can be elaborate (many steps, many tools, they can even call other skills) or dead simple — one of mine is a four-sentence snippet of a prompt I was typing too often, so I turned it into a single command. You can write your own, download other people's or install ready-made packs. They give me the most leverage in my daily work.
How to start: take one prompt you type over and over, and save it as a skill. That's all.
02The status line
The most underrated feature I know. It's a bar at the bottom of the screen that constantly shows which model you're using, what effort level it's running at and — most importantly — how much context you've already used. For example: 27%, that is 274,000 out of the million available tokens.
Why does that matter so much? Because without that number you're working blind. Seeing how much context is left, you know when to hand the work to a new session, when to clear it and when to compact it — before the model starts to lose its way. It was the main reason I moved from a plugin in my editor to working in the terminal. Setting it up takes a moment: you type the status-line command and say what should be on it. You can even show Claude a screenshot and ask it to recreate that layout.
03Routines (scheduled agents)
This is the ability to schedule a task for a specific time — but note, you're not scheduling an ordinary script that runs deterministically the same way every time. You're scheduling an agent. At the set hour your Claude gets to work exactly as if you were talking to it live. That's the difference between a rigid invocation and an intelligent operator that can read the situation.
You set routines with the schedule command, and in the desktop app you can view them in a calendar or a list and choose whether they run locally or in the cloud. A daily review, a recurring report, repeatable research — those are natural candidates.
04Remote control from your phone
The ability to run a local Claude Code session from your phone or a browser. You turn on remote-control mode, get a link, open it on your phone — and the session is fully synced both ways. What you reply from the phone shows up on the computer and vice versa.
I use this when I want to go for a walk but don't want to break the session — a brainstorm or building something keeps going, and I just respond from my pocket. A simple thing that unsticks you from the desk without losing the thread.
05/loop — recurring commands
This is the ability to repeat a single command in the terminal at even intervals. Underneath it uses the cron mechanism — a scheduler that runs a task every set period (every minute, every ten minutes, every hour). You don't even have to type the loop command; you just ask in plain language: "set up a repeating task every twenty minutes."
There's one important detail worth understanding: these loops are tied to the session. Each terminal has its own, and once you close the terminal the loop disappears. By default they also expire after seven days of continuous running. That's healthy — nothing is left running forever behind your back.
The next six — features I reach for constantly
06Subagents
A subagent is a helper launched in the background. It can be invoked automatically or at your request, and several work in parallel. The key difference from agent teams (more on those shortly): subagents don't talk to each other — each reports only to the main session that invoked it. You define them the same way as skills — with a markdown file describing what the agent does, which tools it may use and what skills it carries. I use them very often; it's one of my favorites.
07/rewind — checkpoints
It lets you roll the code and the conversation back to an earlier point. That sounds trivial, but it has a deeper purpose. When Claude does something wrong, you instinctively write "no, that was wrong, do it differently" — and in doing so you add another layer of confusion to the context. It's better to roll back to the moment before the error and start fresh from there. You pick the exact prompt, the exact point in the conversation, and return to it. Less junk in the context, a cleaner result.
08Agent teams
One of my favorite features overall. Unlike subagents, team members talk to each other and can argue. I love building panels and debate boards out of them: I ask for a team where one character is a complete beginner, another a CEO, and others play still further roles — and have them analyze or debate some topic. They run rounds of debate until they reach agreement, and I listen to those different perspectives and form my own view.
It's a costly feature — it burns a lot of tokens — so I reach for it deliberately, for example when I still have plenty of headroom in the current work window and the session is about to end anyway. You have to enable it in the settings first, because it's experimental; then you invoke it in plain, natural language.
09Auto-memory
This is Claude's ability to improve its own memory on its own, without your command. At a set interval — a certain number of sessions, conversations or amount of time — it searches through what you've done together and updates itself. You could say it "dreams" and becomes smarter for it. Check in the memory settings whether you have this feature on; if not, and you want to use it — just switch it on.
10/insights — usage report
It generates a report on the last thirty days of your work in Claude Code and shows what you do a lot of, what you do poorly, which features you could use more often and which skills would be worth building. You get a clear report: the number of messages and sessions, the dates, "what's working," "what's holding you back," quick improvements to try and more ambitious workflows. It's worth looking in on this now and then to see how your way of working is evolving.
11ultra plan — planning in the cloud
It moves planning to the cloud and splits it across many planning agents. I reach for this when I have a large, ambitious task and want to plan it properly before I execute anything. You type the command, decide it should run in the cloud, get a link — and meanwhile you can keep working in the terminal, because the planning tasks run in the background. The plan then comes back to you, or you start carrying it out right away in the browser.
12/goal — work until you hit the goal
You tell Claude what to do, and that it shouldn't stop until it meets a condition. The secret lies in one thing: you have to give an unambiguous definition of done. "Find three very hard puzzles" is weak, because "very hard" is subjective. The more objective the criterion, the better this works.
Contrary to the popular demos, the point isn't for the agent to run for 24 or 48 hours. The value is different: I send off a task with a clear criterion and can walk away — take a break, go to sleep — fairly confident that when I'm back the task will actually be done, because verification is built into it. A good example with a measurable goal: "optimize this page until a given element loads instantly on refresh." That's a real criterion the agent can check itself.
What I deliberately left off the list
A list of the best features is also a list of what I rejected — and that's just as instructive. A few things are very useful, but they don't change my day enough to make the top twelve.
The CLAUDE.md file, the /clear and /compact commands, resuming a conversation, web search — these are the foundation, the "can't move without it," but no one names them as a favorite feature. Voice mode, co-work mode, interactive connectors — great, but I reach for them less often. I value more highly a few things I use as support: the command-line integration with the office suite (my agent then touches spreadsheets, documents and email through one consistent tool), the desktop app for managing routines and reviewing many projects, and prompt caching, which saves a great many tokens while working quietly in the background — you don't even notice it.
The conclusion is consistent with where I started: useful isn't the same as day-changing. You have to separate those two measures, or you'll invest your attention in tools that merely look good.
A rule to close
Don't try to master everything at once. Pick one feature from this list — ideally skills or the status line — and weave it into your work long enough that it stops being a novelty and becomes a habit. Only then add the next one.
The best tool isn't the flashiest, but the one that changes the most in your particular day. Your way of working isn't someone else's — so build your own ranking, not theirs. Start with one prompt you type too often, and turn it into a recipe.