Claude Code: a hands-on review of AI-assisted development

What is Claude Code?

Claude Code is a command-line development assistant built by Anthropic. It allows developers to interact with their codebase in natural language: describing a feature, requesting a refactor, generating tests, fixing a bug, all directly in the terminal.

Unlike assistants embedded in an editor (such as GitHub Copilot), Claude Code operates at whole-project level. It can navigate files, understand the architecture, run commands and propose coherent modifications across multiple files simultaneously.

What works particularly well

Scaffolding and structural code generation

For creating a new component, an API route, a database schema or a project configuration, Claude Code excels. It generates clean, well-structured code that respects the conventions of the existing project. Time saved on these repetitive tasks runs at around 60 to 80%.

Large-scale refactoring

Renaming a variable across 40 files, migrating from one API to another, restructuring a tree: Claude Code handles these operations with high reliability. What used to take a day takes 15 minutes.

Bug fixing with context

Describe a bug in natural language ("the contact form does not send email when the phone field is empty"), and Claude Code identifies the relevant file, understands the logic and proposes a fix. Its ability to navigate the code to find the root cause is a key strength.

Test generation

Asking Claude Code to generate unit tests for a function or module produces good-quality results. The tests cover nominal and edge cases and integrate well with existing frameworks.

Measured productivity gains

Figures vary by task type, developer seniority and project complexity. Here are orders of magnitude observed on real projects run with Claude Code between 2025 and early 2026:

According to an Anthropic study published in 2025, developers using Claude Code report an average productivity gain of 2x to 3x on common coding tasks. This figure is consistent with field observations in French-speaking Switzerland.

An important caveat: the gross gain does not include code-review time, which remains necessary. Including review, the net gain is rather between 40% and 60% across the full development cycle.

Comparison with other AI development assistants

The AI coding assistant market is expanding fast. Here is how Claude Code positions itself versus the main alternatives:

GitHub Copilot runs as autocomplete inside the editor. It excels at line-by-line suggestions but does not understand the project's overall architecture. Claude Code, by contrast, operates at whole-project level and can coordinate changes across multiple files.

Cursor integrates an LLM directly into a fork of VS Code. The experience is fluid for developers working exclusively in that editor. Claude Code is agnostic: it works in any terminal, with any editor, on any project.

Aider is an open-source command-line tool comparable to Claude Code. It supports several models (GPT-4, Claude). Claude Code stands out through its native integration with Anthropic models, its project-memory system (CLAUDE.md files) and its ability to execute system commands.

The choice depends on context. For a team that wants a versatile tool capable of intervening across the full development cycle (from prototyping to production), Claude Code is currently the most complete.

The limits to know

Loss of context on long sessions

This is the most significant limit. On an extended development session (more than 2 hours), Claude Code progressively loses the thread of context. It may propose modifications that are inconsistent with what was done at the start of the session, or forget decisions taken earlier.

This phenomenon is linked to the limited size of LLM context windows. The longer the session, the more the model "forgets" the early exchanges.

Complex architectures

For structuring architectural choices (microservices vs monolith, choice of patterns, design of distributed systems), Claude Code remains an assistant, not an architect. It proposes solutions that work, but not always those best suited to the project's specific context.

False confidence

Generated code looks professional and often compiles on the first try. That appearance of quality breeds false confidence. Without careful review, subtle bugs, security flaws or sub-optimal choices slip through.

Cost and token consumption

Claude Code consumes tokens at every interaction. On a complex project with many files, a single query can consume tens of thousands of tokens of context. For a team of 5 developers using the tool intensively, monthly budget can reach CHF 500.– to CHF 1'500.– depending on the chosen plan. This cost should be weighed against the productivity gain, but it deserves to be budgeted from the start.

Concrete examples of projects built with Claude Code

Next.js showcase site (MCVA Consulting)

The site you are reading was built with the help of Claude Code. Component structuring, dynamic page generation, headless CMS configuration, CI/CD pipeline setup: Claude Code accelerated each step. Total development time was reduced by about 50% compared to a classic approach.

SEO/GEO audit platform

For a project to build an audit platform combining traditional SEO analysis and Generative Engine Optimization (GEO), Claude Code was used to generate database schemas, API routes, interface components and scoring algorithms. The tool's ability to maintain consistency between frontend and backend on a multi-module project was decisive.

Technical migration of a legacy application

On a mission to modernise a PHP application towards a TypeScript/Node.js stack, Claude Code automated a large part of the rewrite. The tool analysed the existing source code, identified recurring patterns and generated modern equivalents. Migration time was divided by three.

Best practices for teams

Short sessions, targeted objectives

The most impactful practice: work in 30 to 45-minute sessions with a precise goal per session. "Build the dashboard component", "Fix bug #42", "Add tests for the payment module". Then close the session and open a new one.

This approach circumvents context loss and produces more reliable results.

Systematic code review

Every change proposed by Claude Code should be reviewed like any pull request from a junior colleague. Check the logic, security, performance and consistency with the existing architecture.

Document effective prompts

When a prompt produces a good result, document it. Build a shared prompt library within the team. It is a knowledge asset that grows over time.

Combine with vibe coding for prototypes

For rapid prototyping with vibe coding, Claude Code paired with this approach is very effective. Describe the desired application, let Claude Code generate the prototype, test, iterate. For production, switch back to a rigorous development mode.

Implications for Swiss businesses

A competitive advantage for SMEs

In Switzerland, the hourly rate of a senior developer ranges from CHF 150.– to CHF 250.–. A tool that reduces development time by 50% represents substantial savings, even accounting for the licence cost. For an SME hesitating between SaaS and custom development, Claude Code changes the equation: a small team of 2 to 3 people can produce as much as a team of 5.

Compliance and data sovereignty

Swiss companies are rightly attentive to data sovereignty. With Claude Code, the source code is sent to Anthropic's servers for processing. For projects subject to strict regulatory constraints (banking, health, public administration), this question must be evaluated case by case. Anthropic offers contractual commitments on data non-retention, but each organisation must conduct its own risk analysis.

Train rather than replace

The aim is not to replace developers with AI, but to make them more effective. The companies that get the most out of Claude Code are those that invest in training their teams: understanding the tool's capabilities and limits, mastering prompt engineering applied to code, integrating AI into existing workflows without breaking quality processes.

Adoption in French-speaking Switzerland

Swiss companies are starting to adopt Claude Code and similar tools. Initial feedback is positive, provided best practices are followed and teams are properly trained.

The most advanced sectors in adoption: digital services companies, fintech start-ups and large companies' IT departments. But the tool is relevant for any organisation that develops or maintains code.

The question is no longer "should we adopt AI-assisted development?" but "how do we adopt it effectively?". That is precisely the question MCVA Consulting helps Swiss companies answer.

Summary

• Claude Code excels in scaffolding, refactoring, bug fixing and test generation.
• Its main limit: context loss on long sessions (over 2 h).
• Key best practice: short sessions (30-45 min) with a single, precise goal.
• Every change must undergo a systematic human code review.
• Contact MCVA Consulting to support the adoption of AI-assisted development in your company.

Frequently asked questions

Can Claude Code replace a developer?

No. Claude Code is an accelerator, not a replacement. It automates repetitive, structural tasks, but architectural decisions, understanding of business needs, code review and validation remain human responsibilities. A developer equipped with Claude Code ships more, and ships better. Claude Code without a developer produces code that compiles but does not necessarily solve the right problem.

What is the real cost for a development team in Switzerland?

Anthropic's Max plan (which includes Claude Code) is around USD 200.– per month per user. For a team of 3 developers, that is around USD 600.– per month (around CHF 530.–). Comparing with the estimated productivity gain (40 to 60%), ROI is generally reached in the first month. For intensive API use, costs can be higher and must be tracked closely.

How do you integrate Claude Code into an existing workflow without disrupting the team?

The key is gradual adoption. Start with a non-critical pilot project. Identify one or two motivated developers to test the tool for 2-3 weeks. Document the gains, limits and best practices observed. Then progressively extend it to the whole team, supporting each step with appropriate training.

Want to assess how Claude Code and AI-assisted development can accelerate your projects? Contact MCVA Consulting for a tailored diagnosis and adoption plan suited to your context.