Some useful Claude Code tips

After some time working with Claude Code I've adquired some habits that improve both the speed, clutter, and overall efficiency of the tool. Here are some of them.

Tip #1: CLAUDE.MD best practices

CLAUDE.md is a Markdown file placed at the root of your project that Claude Code automatically reads at the start of every session (yes, you can set a global one). It acts as persistent instructions — your project's "constitution." But it can also reduce your AI available tokens.

Here, I will mostly repeat the same I said on Vibe coding a blog from scratch: how it actually went

Give it two sections:

1. Proyect instruccions: keep it short. Write what goes into scope and what's not. Tech stack, project structure, ask before do, etc.
2. Add this exact text:

Direct and casual tone.
No bullet points unless I ask for.
Never use filler words like "certainly", "absolutely", "great question" or "I'd be happy to".
Read existing files before writing. Don't re-read unless changed.
Thorough in reasoning, concise in output.
No sycophantic openers or closing fluff.
No emojis or em-dashes.
Do not guess APIs, versions, flags, commit SHAs, or package names. Verify by reading code or docs before asserting.

Tip #2: how to fix slop code

To prevent AI from being excessively compliant and overlooking critical errors in its own output, you must assign it the role of a severe, skeptical, and "bad-mood" code reviewer who adheres strictly to the rulebook without granting any exceptions.

Instead of simply asking "what do you think of this code?", you must instruct it to assume the code was written by another model or a careless developer, and that its sole mission is to uncover flaws, redundant code, unnecessary logic, or "slop," forcing it to adopt a critical stance.

It will aggressively point out any suspicious patterns, missing optimizations, or violations of best practices, which, paradoxically, makes the AI significantly more accurate and effective at detecting its own mistakes because it breaks its natural tendency toward automatic validation and politeness.

This will transform the AI from an accommodating collaborator into an implacable auditor that prioritizes technical correctness over harmony.

Tip #3: Edit a message instead of adding more on top

When Claude Code produces something close but not quite right, resist the urge to send a chain of corrective messages on top.

Each new message adds to the context, and contradictory instructions scattered across multiple messages degrade output quality.

Instead, edit your original prompt to incorporate the correction directly. This forces you to articulate a cleaner, more precise instruction and keeps the conversation thread coherent.

Claude then processes one unified directive rather than trying to reconcile "do X" from message one with "actually, do Y instead" from message four.

Tip #4: /clear + /compact

These are two essential context-management commands that will help you keep Claude tokens in check.

/clear

/clear wipes the entire conversation history and starts fresh — use it when you're switching tasks or when the context has become large (+15 messages). Claude will save the state of the conversation to disk, and you can get back to it later by issuing /resume.

/compact

When you think your context window is too large (+15 messages) but you don't want to give up all the context you can use /compact to reduce it and save tokens. /compact condenses the existing conversation into a summarized form, preserving key decisions and context while freeing up token space.

The power move is combining them strategically: use /compact when the conversation is still useful but getting long, and /clear when you need a hard reset. This prevents expensive token usage and the degraded performance that comes from bloated context windows filled with noise.

Tip #5: Stay in control: IDEAS.MD

As explained in Vibe coding a blog from scratch: how it actually went having a "living backlog" of ideas and decisions is a key part of the project's success.

It allows you to stay in control of the project's direction and also always intruct concice tasks and the order in which they should be done.

While CLAUDE.md handles project rules and conventions, IDEAS.md is a scratchpad for architectural thoughts, feature proposals, and open questions.

You can hold a bullet-point list or a list of detailed sections representing the ideas you have in mind. One is shorter, the other is more detailed.

You can tell claude to propose ideas and incorporate them into IDEAS.md or manually add them yourself.

It also serves as a guide for the AI to follow. Claude will read and hold this information in memory, so it will not do some idea that may affect others without telling you.

Tip #6: Plan Mode

Plan Mode tells Claude to think before acting.

Instead of immediately writing code, Claude first produces a detailed plan: what files to modify, what approach to take, and what the execution order should be. You review and approve (or adjust) the plan before any code changes happen.

This is critical for complex tasks where a wrong assumption early on cascades into wasted effort.

Tip #7: Text guards are not enough

Be careful to not overlook the fact that LLMs will not follow your instructions exactly (or even do the opposite).

You can set some instrucctions as "text guards" and tell the AI to always follow these rules.

BUT mind that this is not a guarantee that the AI will follow them

There's no such a thing as Asimov's three laws of robotics when speaking about LLMs

Final thoughts

This is just a small list of tips that I've found useful and proven right every time so far. Some of them allows you to save token usage, and others improve the quality of the output or the overall experience with Claude Code.