---
title: "Vibe coding with Claude: five practical tips you need to try out"
description: "Vibe coding with Claude: five practical tips with step-by-step instructions—from Plan Mode and CLAUDE.md files to Artifacts."
url: https://johannesbecht.com/blog/vibe-coding-with-claude
date: 2026-07-25
modified: 2026-07-25
author: "Johannes Becht"
image: https://johannesbecht.com/wp-content/uploads/2026/07/vibe_coding_mit_claude_illustration-1-scaled.png
categories: ["AI Blog"]
type: post
lang: en
---

# Vibe coding with Claude: five practical tips you need to try out

I still remember the moment when I first truly understood what “vibe coding” meant. I was sitting at my laptop one evening with a vague idea for a small tool in my head, and I simply started describing what I wanted—without writing a single line of code myself. Ten minutes later, I had a working prototype in front of me. No Stack Overflow tab, no framework documentation, and no worrying about syntax. Just me, my idea, and Claude taking care of the rest.

That is what vibe coding with Claude is all about, and it feels incredible to be able to turn ideas into reality so much more easily—and quickly.

That feeling is exactly why so many people, experienced developers and complete beginners alike, are starting to work in an entirely different way. But simply prompting without a plan is not the same as building well with [Claude](https://claude.ai/). A few techniques can make the difference between “well, that kind of worked” and “wow, that genuinely saved me time.” Those are exactly what I will show you here, with every step explained so you can try them yourself.

Vibe Coding with Claude – are you ready?

## Vibe Coding with Claude: Why So Much Hype??

The term “vibe coding” itself is still relatively new. [AI](https://johannesbecht.com/blog/category/ai) researcher Andrej Karpathy coined it in a post in early 2025. Paraphrased, the idea is that you completely give in to the vibe, almost forget that code exists behind the scenes, and simply describe what you want to happen. What began as an inside joke became an established part of developers’ vocabulary within a few months—and has since even found its way into dictionaries.

There are concrete reasons why Claude is mentioned so often in this context. This is no longer just about completing individual lines of code, as you might expect from traditional coding assistants. Claude—particularly through Claude Code, its agentic command-line tool—works in a genuine loop. It independently searches your project, develops a plan, selects appropriate tools, modifies files, runs commands and tests, reviews the results, and corrects itself when necessary. It can often work across multiple files without requiring you to explain every individual step.

What makes vibe coding with Claude even more interesting is that its autonomy is not completely unrestricted. Features such as an explicit Plan Mode, project-specific memory files, and clearly defined subtasks allow you to control exactly how much freedom Claude has—from “show me a plan first” to “go ahead and do it.” This combination of speed and control is why vibe coding with Claude does not feel like flying blind to many users. Instead, it feels like working with a highly independent but manageable developer colleague. That is precisely where the following five tips come in.

## Tip 1: Create a CLAUDE.md—the Memory for Your Project

The most common mistake in vibe coding is explaining the same things to Claude at the beginning of every new session: how the project is structured, which conventions apply, and how the code should be tested. That takes time and leads to inconsistencies. The solution is a simple Markdown file that Claude Code automatically reads at the beginning of every session.

Here is how to set it up:

1. Create a file named exactly `CLAUDE.md` in your project’s root directory.
2. At the top, briefly explain what the project is about, including its name, purpose, and a one- or two-sentence description.
3. List the technology stack, such as “React, TypeScript, Vite, Tailwind.”
4. Add the most important commands Claude would otherwise have to guess: the build, test, and start commands, such as `npm run build`, `npm test`, and `npm run dev`.
5. Document your coding conventions—for example, “no `any` in TypeScript,” “no default exports,” and your preferred commit format, such as `feat/fix/chore(scope): description`.
6. Add brief notes about the folder structure or architectural decisions that might not be immediately obvious.
7. Save the file—and you are done. From the next session onward, Claude Code will automatically read it without you having to mention it.
8. Keep the file up to date. Whenever your conventions change, update it immediately instead of putting it off until later.

By the way, if you work directly in Claude.ai rather than in the terminal, Projects and their associated project knowledge follow a very similar principle. You can store the relevant context once instead of entering it again and again.

## Tip 2: Activate Plan Mode Before Making Major Changes

The biggest frustration when vibe coding with Claude usually happens like this: Claude eagerly changes several files, the result is “almost right,” and you then spend several messages correcting it. Plan Mode is designed to prevent exactly that. It makes Claude present a plan first without touching your code until you approve it.

Here is how to activate and use it:

1. Start Claude Code in your project’s terminal with the command `claude`.
2. Press `Shift+Tab` twice to switch to Plan Mode. A confirmation will appear in the footer showing that Plan Mode is active.
3. Alternatively, type `/plan` at the beginning of your next message if you only want Claude to create a plan for that particular prompt.
4. Describe your task as precisely as possible—for example: “Add a rate limiter to all API endpoints that returns a 429 response with a Retry-After header.”
5. Claude will inspect your project in read-only mode without modifying a single file.
6. It will present you with a step-by-step plan. Read it carefully instead of merely skimming it.
7. Approve, modify, or reject the plan if something does not look right.
8. Once you approve it, Claude will automatically leave Plan Mode and implement the changes.
9. If you want to start an entire session in Plan Mode—for example, when working with an unfamiliar repository or security-critical code—launch it with `claude --permission-mode plan`.

A practical rule of thumb: Plan Mode is almost always worth using when a change affects three or more files, involves database schemas, or touches code you do not yet know well. You can safely skip it for a single minor correction, where it would usually be overkill.h der Plan-Modus fast immer. Bei einer einzelnen kleinen Korrektur darfst du ihn getrost weglassen – dafür ist er overkill.

## Tip 3: Describe the Vibe, Not Just the Syntax—and Work Iteratively with Images

The name “vibe coding” is no coincidence. You often get the best results when you do not write your prompt like a technical specification. This is also true when vibe coding with Claude. So what you do is describe the desired feeling, purpose, or appearance, and then refine the result over several rounds rather than trying to define everything perfectly in the first prompt.

Here is how to do it:

1. Begin with a broad description of the desired result. “A playful, friendly landing page for a children’s toy” is often more useful than twenty technical requirements at once.
2. If available, add a reference image, screenshot, or mockup directly to the chat. Claude can process images as input and use them as visual guidance.
3. Ask Claude to create an initial version and review the result—for example, through the live Artifact preview in Claude.ai.
4. Give feedback in natural language, just as you would when speaking to another person: “add more white space,” “the button feels too bulky,” or “use a warmer color.”
5. Let Claude iterate instead of prescribing every detail yourself. Set the general direction, evaluate the result, and refine it. That is the actual vibe-coding loop.
6. If you are unsure which direction works best, ask for alternatives—for example: “Show me two different versions of this.”
7. Once the overall direction feels right, become more technically specific. You can then address accessibility, performance, and edge cases in separate follow-up prompts.

## Tip 4: Divide Complex Tasks Among Subagents

As soon as a task becomes larger—for example, “explore the codebase first” followed by “build a new feature”—it can help to divide the work among clearly defined subagents. A subagent is essentially a specialized instance of Claude with its own restricted role. It could, for example, explore a project in read-only mode without being able to change anything accidentally.

Here is how to create one:

1. Create a `.claude/agents/` folder inside your project if it does not already exist.
2. Create a new Markdown file inside it, such as `explorer.md`.
3. Add a YAML frontmatter block at the top containing at least `name` and `description`, such as `name: explorer` and `description: Read-only exploration of the codebase`.
4. Restrict the permitted tools in the frontmatter—for example, use only `tools: Read, Grep, Glob` for a subagent that must not be able to modify anything.
5. In the main body of the file, clearly describe what the subagent should do—for example: “Map the repository structure and summarize the most important entry points.”
6. Let the main Claude instance use this subagent for the appropriate subtask while you continue thinking about the actual feature.
7. Use different subagents for different purposes—for example, one read-only explorer and a separate reviewer that checks changes against your team’s standards.
8. The main agent brings the subagents’ findings together and uses them to plan the next steps.

The advantage is that each subagent remains clearly limited to its assigned role. This can be particularly reassuring when working with larger or security-sensitive codebases.

## Tip 5: Use Artifacts in Claude.ai to Create Prototypes Without Any Local Setup

Not every vibe-coding session requires a terminal, repository, or development environment. For quick prototypes, small interactive tools, or diagrams, an ordinary conversation on Claude.ai is often enough—thanks to Artifacts.

Here is how it works:

1. Open a new chat on Claude.ai.
2. Describe what you want to build, such as a small interactive tool, a sample website, or a diagram.
3. Claude will automatically create an Artifact—for example, HTML or React code—and display a working live preview alongside the conversation.
4. Test the preview directly in the chat by clicking buttons, filling in form fields, and trying out the interactions.
5. Describe any desired changes in your next message, such as “make the buttons bigger” or “add a third field.” Claude will update the Artifact accordingly.
6. Repeat steps four and five until you are satisfied with the result.
7. Once you like it, you can view the code and reuse it in your own project.

The best part is that you do not need to install anything locally. This makes Artifacts ideal for turning an idea into something visible within minutes—before you even decide whether it is worth developing “properly.”s wert ist, “richtig” gebaut zu werden.

## Conclusion: Vibe Coding with Claude

Vibe coding with Claude can sometimes feel almost too easy—and that is both its greatest strength and its biggest trap. For prototypes, internal tools, experiments, or the first draft of an idea, this relaxed, conversational workflow is incredibly fast. But “it works” is not the same as “it is clean, secure, and maintainable.” Critics of the movement rightly point out that good software still requires care, even when the vibes are right.

That is precisely why these five tips work so well together. A `CLAUDE.md` gives Claude persistent context, Plan Mode prevents expensive surprises, iterative work with images helps you find the right direction more quickly, subagents keep larger tasks clearly separated, and Artifacts can transform a vague idea into something tangible within minutes.

Try whichever of these five techniques catches your attention first the next time you want to start building something with Claude. I am fairly certain that once you have used them, you will not want to go back.