You're using Claude Code wrong. 5 habits that change that.
Claude Code is my standard tool by now. Every day. For code, for writing, for entire workflow designs.
In the first months I lost more time with it than I saved. Correction loops. Outputs that didn't fit. Sessions that got worse the longer they ran.
Then I understood the mechanics.
It's not about tricks. It's about one single mindset, and five habits that put that mindset to work.
The mindset first: you treat Claude like a machine. That's exactly the problem.
As long as you use Claude Code like a tool, instruction in, output out, you get tool output. As soon as you treat Claude like a sparring partner, it gets a lot better. So do you.
Here are the five habits that make this possible in practice.
The most common mistake when working with Claude Code: you give an instruction, Claude gets going, and you spend the next hour typing in corrections.
Plan mode flips that around. Claude may read, analyze, suggest, but not change anything.
In this mode the most important part of the whole session happens: you understand each other.
You give feedback. Claude adjusts the plan. You give feedback again. For as long as it takes until what's in your head is cleanly laid out in the plan. Only then does the building start.
Sounds slow. It isn't. You save yourself the double and triple corrections later.
This isn't a software engineering method. It's strategy.
Ask before you build:
One single sentence I send before every bigger task: "Ask me follow-up questions until you're completely sure what I want. Only then start."
From that moment on, Claude-guesses-and-you-correct is over. Instead, a list of concrete questions comes back.
Which colors? Which sections? Where does the call to action go? What should this edge case look like?
Sounds trivial. But it closes the biggest hole in the workflow: assumptions nobody said out loud.
Discuss instead of command:
Compare two prompts. First the obvious one: "Build me a contact form with name, email, phone number and message field."
Then the better one: "What should a contact form look like so that as many people as possible fill it out?"
The first prompt gets you a standard form. The second gets you two variants, a recommendation and a note that you forgot to think mobile-friendly in the first version.
Claude isn't there to fulfill your specs. Claude is there to improve them.
Anyone who doesn't get that has hired a hugely expensive intern who only does what he's told. And nothing more.
Keep CLAUDE.md up to date:
Every session in Claude Code starts from zero. Claude no longer knows what was discussed in an earlier session. Unless it's in a CLAUDE.md.
The CLAUDE.md is the project memory. It loads automatically in every new session. Whatever you put in there, Claude knows right away next time.
What belongs in it: everything Claude keeps getting wrong. Conventions nobody else documents. Quirks of your stack. Tone rules, if you write copy.
What doesn't belong in it: style instructions that change every three days. The file gets loaded every time. It costs tokens.
Rule of thumb: under 200 lines. Anything beyond that you move into other files and only reference.
A well-kept CLAUDE.md is the difference between starting-from-zero-every-time and a system that grows with you.
One task per chat:
The longer a session runs, the worse Claude gets.
There's a technical reason for that: every language model has a limited context window, and for Opus 4.7 it's one million tokens, so roughly 750,000 words. Sounds like a lot. It is. But the fuller that window gets, the more the quality suffers. The effect is called Context Rot.
In a cluttered chat, Claude digs through 200,000 words of old instructions for what you need now. Half of it is irrelevant. The other half contradicts itself.
The fix is simple: New task? New chat. Need context from the old chat? /compact writes a summary, keeps the essence, throws out the rest.
Once you've taken that in, you'll never again wonder why Claude suddenly turns strange in the third hour.
The rest, in two sentences:
Worktrees, subagents, custom skills, hooks, voice input, screenshots as a brief, Ultraathink mode for hard problems, all solid. All useful once you have the workflow down.
Without the five above, they're only features. With them, they become a lever.
What sticks:
Most people ask: "What can AI do?"
That's the wrong question.
The right question is: "How do I talk to AI?"
Once you understand that, you stop collecting tools. And start building a style.
Same as with brands.
FAQ
What is plan mode in Claude Code?
Plan mode is a state in which Claude Code may read files and make suggestions but can't change anything. You give feedback until the plan shows exactly what you want. Only then does the implementation start. That prevents expensive correction loops afterward.
What exactly is the CLAUDE.md file?
The CLAUDE.md is a text file that works as the project memory for Claude Code. It loads automatically in every new session and holds a project's conventions, recurring mistakes and quirks. Without it, every session starts from zero, because Claude doesn't remember anything from earlier sessions.
What does Context Rot mean in Claude Code?
Context Rot is the loss of quality that sets in when a context window gets too full. The more old, partly contradictory instructions pile up in a long session, the harder it is for the model to find the relevant information. The result is worse and less fitting answers the longer the chat runs.
How big is the context window of Opus 4.7?
According to the article, Opus 4.7 has a context window of one million tokens, which is roughly 750,000 words. Despite that size, answer quality suffers once the window fills up with old instructions. That's why a new chat is the recommended move for every new task.
How long should a CLAUDE.md be at most?
The rule of thumb is under 200 lines. Anything beyond that should go into separate files that are only referenced. The reason: the file loads in full in every session and uses up tokens doing so.