"Is Claude Code just about prompts? Do settings even matter?"
"I installed it, but it wasn't as smart as I expected."
"CLAUDE.md seems like a hassle. It works without it anyway."
All of these were wrong.
Actually, Claude Code's performance improves dramatically just by changing settings
In this article, I will introduce 10 settings that you can try today, ranging from those who have just installed Claude Code to those who are already using it but are dissatisfied with its accuracy. None of them are difficult. However, whether you do them or not will make Claude's output feel like a completely different thing, so please give them a try.
Before we get into the main topic, here is an announcement.
We will be holding the AutoThreadsMonetize seminar for 4 days only from 6/5 (Fri) to 6/8 (Mon)!
I will reveal how even a complete amateur with no SNS experience can grow Threads from a fresh account and monetize it, so if you wish to participate, please make a reservation here.
→https://utage-system.com/p/1fPuHlixHYbu?ftid=smxCqKpT9Fe6
Now, let's dive into the 10 settings that dramatically improve Claude Code's performance.
① Automatically generate a CLAUDE.md draft with /init
The first thing to do is simple. Run /init. That's it.
When you run /init, Claude analyzes the entire codebase and automatically creates a draft of CLAUDE.md, including the directory structure and project conventions. It's obvious which is faster: driving a car without setting a navigator, or driving after entering the destination.
However, this is important: do not trust the results of /init blindly.
Information Claude can find on its own is different from your implicit knowledge.
"Don't touch this folder."
"The purpose of each folder."
/init is just a draft. A human must review it and add implicit knowledge. This one extra step makes the subsequent accuracy of Claude significantly different.
② Trim CLAUDE.md to under 200 lines
The CLAUDE.md generated by /init is often too long as it is.
Anthropic's official documentation also recommends "keeping CLAUDE.md concise." The reason is simple: the contents of CLAUDE.md are loaded into the context every time a session starts. If you stuff it with long manuals or temporary notes, it will squeeze the context and actually degrade performance.
You should only write "facts you want Claude to know every time."
What to write varies depending on each person's use case. Naturally, the rules required for information dissemination and tool development are different.
For now, it's good to write the directory structure, the role of each folder, and rules you absolutely want it to follow.
Keep three things in mind:
- Short
- Specific
- Is it a rule needed every time?
If there are too many rules to fit in 200 lines, you can split them into the .claude/rules/ directory.
You can specify the scope of application for each path, so it can be loaded only when necessary, eliminating the waste of squeezing the context during unrelated work.
③ Organize allow/deny permissions in .claude/settings.json
Claude Code can perform operations that affect the actual development environment, such as file editing and Bash commands. That's why permission design is necessary.
Set allow and deny rules in .claude/settings.json. The evaluation order is deny → allow, with deny taking top priority.
For example, set commands that forcibly delete files or folders, like rm -rf, to deny to prevent them from being executed without permission.
For those who want to minimize risk but find it tedious to be asked for permission every time, simply setting allow/deny will improve work efficiency.
Prevent through "mechanisms," not "requests." This is true security design.
In team development, commit .claude/settings.json to the repository to share it with everyone. If you separate individual settings into settings.local.json, accidents won't happen.
④ Clean up old memories with /memory
Claude Code has a mechanism called "Auto memory" that carries knowledge across sessions. It learns and saves from user corrections and preferences. It's convenient, but it can also be a bit of a nuisance.
There is a risk that old assumptions or incorrect learning will remain left behind.
Using the /memory command, you can check what is currently loaded. Check everything and delete any old instructions that are no longer used.
When Claude repeats the same mistake, suspect the memory first before changing the prompt. Surprisingly, this often solves the problem.
AI memory becomes a liability if left alone. Just putting a monthly memory inventory into your routine will change the stability of accuracy.
⑤ Start complex changes with Plan mode
If you let Claude implement things suddenly, it might run at high speed in the wrong direction.
Especially for complex changes, first separate investigation and planning with Plan mode. Anthropic's official best practices also recommend the flow of "Exploration → Planning → Implementation → Commit."
Usage is simple.
Just press Shift+Tab to switch to Plan mode.
In Plan mode, Claude reads files and investigates without performing actual edits. You can review it before it heads in a strange direction.
⑥ Have it write implementation plans like requirement definitions and tests first
Even if you decide on a Plan, if you let Claude Code implement it without a requirement definition document, it might forget the specifications halfway through and end up with something different from the plan.
Also, by writing completion conditions like tests in the implementation plan beforehand, Claude Code itself can verify whether it has completed the task appropriately.
By following a fixed flow from plan to plan creation and review every time, the accuracy of implementation will improve significantly.
That said, I think it's difficult for non-engineers to do this suddenly, so for those people, the Superpowers plugin is recommended. Since appropriate flows are defined from the start, implementation will be completed with high accuracy just by following the instructions.
If you haven't installed it yet, please try the plugin.
⑦ Automate formats and prohibited operations with Hooks
CLAUDE.md is a "request." Hooks are "compulsory." This difference is important.
Hooks are shell commands that are automatically executed at specific timings in Claude Code. They operate deterministically without relying on the LLM's judgment. In other words, you can create a structure where Claude cannot make a "careless mistake."
Specifically, use them like this:
- Block
rm -rforgit push --forcebefore execution → Prevent fatal accidents - Detect and stop edits to protected files (
.envor production settings) - Notify the desktop when Claude is waiting for judgment
"Rules you want it to follow" are written in CLAUDE.md, but this is strictly on a request basis.
Even if you say don't read the .env file, things like "This is my mistake. I accidentally read it" happen all the time.
Therefore, using Hooks for processes you want to execute forcibly or prohibit is the correct answer.
This can prevent careless mistakes and omissions.
⑧ Delegate research and reviews to subagents
Claude Code has a mechanism to use subagents that run in parallel, separate from the main agent.
Subagents have the advantage of being able to proceed with work in parallel, but even more significant is that they can separate contexts.
This is effective for research and reviews.
For example, when conducting a large amount of research, even if all you need is a summary of the research results, doing the research with the main agent will clutter the context with unnecessary information.
Therefore, by having a subagent do the research and report the results to the main agent, you can put only the necessary information into the main context.
Also, when reviewing written code, a subagent with a fresh context different from the one that wrote the code is less likely to overlook things.
It's the same as with humans, where oversights are more likely to occur if the creator and the reviewer are the same person.
⑨ Turn repetitive tasks into skills
If you are writing the same prompt every time, you should turn it into a skill.
In Claude Code, by placing a SKILL.md file in the .claude/skills/ directory, you can add reusable procedures and knowledge to Claude's toolkit. Once created, Claude will automatically refer to it when necessary.
If you don't know what to turn into a skill, instruct Claude Code to "Check this directory content and past logs and suggest things that should be turned into skills," and it will make suggestions.
To clarify the difference from CLAUDE.md:
CLAUDE.md is "prior knowledge that is always loaded."
Skills are "manuals that are called when needed."
The point when creating skills is progressive disclosure.
Only the name and description are loaded every time, and the rest is read only when the conditions written in the description are met.
If you specify files to be read for each condition further ahead, it can read only the necessary files, improving performance.
Claude Code has a skill-creator skill, so using that will create a skill according to the format, but it doesn't hurt to know the concept of progressive disclosure.
⑩ Set up CLI tools first
Claude Code can utilize CLI tools when interacting with external services. gh for GitHub, aws or gcloud for cloud, sentry-cli for monitoring, etc.
For example, if you use GitHub integration, finish gh auth login beforehand.
Rather than manually pasting external information into the AI, it's more stable to have it directly retrieve necessary information via CLI. Creating a state where Claude can go get the latest information itself reduces the risk of guessing based on old information.
Even for services other than GitHub, if there is a CLI for the service you are using, just authenticate it beforehand. It's a small effort, but the effect is large when used daily.
MCP is also convenient, but it squeezes the context, so for things that can be handled by CLI, CLI is recommended. Well, this depends on the use case, but you should avoid putting in a lot of MCPs you don't use, as it leads to context pollution.
Summary
The accuracy of Claude Code is determined by the depth of the setup, not the cleverness of the prompt. Arrange the mechanism for passing information, design permissions, and specify verification methods. If these are in place, the output will be different even with the same Claude Code.
As a note of caution, please don't let setting things up perfectly become the goal itself. Spending 3 days polishing CLAUDE.md while the essential work isn't progressing is putting the cart before the horse. Remember that settings are just a means to an end.
I usually post the latest AI information and monetization methods using AI on X, so please follow me.
Also, we will be holding the AutoThreadsMonetize seminar for 4 days only from 6/5 (Fri) to 6/8 (Mon)!
**I will reveal how even a complete amateur with no SNS experience can grow Threads from a fresh account and monetize it, so if you wish to participate, please make a reservation here.
**※ 55 major benefits including the Claude Code and Codex textbook, from introduction to basic operations, are also being distributed on the official LINE.
→https://webiz-assist.com/DYu9CAKZMguw**](https://t.co/aa0yEi6R9N)
If you found this article helpful, I would be happy if you could post your thoughts in a quote post.
