My Journey with AI Tools for Coding
Lee Robinson (Cursor) recently wrote a thought-provoking article about Personal Software - the idea that AI tools now let people build tailored solutions for their specific problems. It's something I strongly relate to. Instead of adapting to generic tools, you can now describe your needs and get help building something that fits exactly.
This made me reflect on my own journey with AI-assisted coding - from pasting error messages into ChatGPT to having an assistant that understands my entire codebase.
Web Interfaces
My first interactions with LLMs for coding were mostly about debugging. An error appeared, I copied the stack trace, pasted it into a web interface, and got back an explanation and potential solution.
Limitations:
- Poor context - the AI only saw what I pasted
- Larger files became difficult to handle
- Lots of manual copying back and forth
- No understanding of project structure
- Outdated knowledge - training data didn't keep up with new versions of frameworks like Next.js and React
- Had to manually paste documentation to get accurate answers
It was useful for isolated problems, but limiting for everything else.
IDE Plugins
When tools like Supermaven (now acquired by Cursor), Copilot, and Cursor started appearing, the dynamic changed. Suddenly the AI was where I worked. Autocompletion that understood the code around the cursor, ability to chat without leaving the editor.
It was a step in the right direction - but still limited to the file I had open.
Specialized Tools
In parallel, a category of specialized web tools has emerged. v0 and Lovable focus on rapid prototyping - from idea to working UI in minutes. Want a booking page for your business? Lovable builds the frontend and forms.
Chef takes it a step further and generates the entire backend system too - database, APIs, and logic.
Strengths:
- Quick visualization of ideas
- Focus on design and components
- Easy deploy or copy-paste to existing projects
- Lower barrier for non-developers
They fill a different role than general code assistants - more "show me what it could look like" than "help me build the entire system".
CLI Tools
The next step in the evolution is CLI-based AI assistants. There are several options - Claude Code, Codex CLI, and Gemini CLI. I've chosen Claude Code as my primary tool.
Unlike IDE plugins, these tools work directly in the terminal with access to the entire project context.
What makes the difference:
- Project understanding - the AI can read files, follow imports, and build a model of the entire codebase
- Direct changes - instead of suggesting code to copy, the tool can edit files directly
- Tool integration - ability to run build, tests, linting, and other CLI commands
- Activities outside the file - create new files, manage git, search documentation
But it doesn't stop there. Claude Code comes with an entire ecosystem of features that extend AI support:
- CLAUDE.md - Project context that defines architecture, commands, and conventions
- Hooks - Automated actions at specific events
- MCP (Model Context Protocol) - Integrations with external services like documentation and APIs
- Skills - Predefined prompts for recurring tasks
- Plugins - Extensibility for specific needs
Together, this creates a workflow where the AI doesn't just answer questions but can actively help plan features, conduct security reviews, discuss architecture, and research solutions.
How I Use It Today
Almost all code I write today is generated with AI assistance. It's not just about getting help with syntax - it's the entire workflow:
- Feature planning - Discuss implementation before writing a single line of code
- Architecture discussions - Get input on structural decisions
- Security reviews - Automated scanning for potential vulnerabilities
- Research - Find the right libraries, understand APIs, compare solutions
- Documentation - Generate and maintain technical documentation
Homelab Documentation and Automation
My homelab consists of multiple servers and services - everything from reverse proxy and DNS to monitoring and game servers. With the AI assistant, I've been able to:
- Inventory servers and read logs via Bash
- Document the entire infrastructure as Infrastructure as Code
- Integrate with third-party services via MCP (Cloudflare, GitHub)
- Apply a security-focused approach throughout
The result is a Git repo with complete documentation that the AI can reference for future changes.
Dotfiles with Stow
My dotfiles are now managed with GNU Stow for symlinks. Claude Code helped restructure configurations and set up a maintainable system.
Screenshot Tool
The screenshot tools I found either had poor UX or were paid services. So I built my own with React and Canvas:
- Anonymization/blur of sensitive areas
- Presets for backgrounds and sizes
- Export to PNG or WebP
A simple tool that does one thing well - exactly what I needed.
Web Scraping
Scripts to extract data from websites and convert to usable formats. Instead of manually copying, I can now run a script and get structured data.
Features in Published Applications
Amealyze - AI-driven security checks and code review on pull requests. Claude Code helped implement and fine-tune it.
Matematiskt.se - Set up events and dashboards for web analytics in PostHog, entirely via PostHog's MCP integration. No manual configuration in the UI needed.
Fast Response to Security Incidents
When CVE-2025-29927 (Next.js middleware bypass) was published, I needed to quickly verify that my projects were secure. One prompt to Claude Code that:
- Scanned through all my projects
- Created necessary changes
- Tested that everything worked
- Managed GitHub issues
- Published updates
What could have taken an entire day was done in under an hour.
Tools
| Tool | Description |
|---|---|
| Claude Code | CLI-based AI assistant with project understanding and tool integration |
| Supermaven | Fast code completion in the editor (sunsetting 2025, acquired by Cursor) |
| Context7 | MCP server for library documentation |
| Cursor | AI-first IDE built on VS Code |
The Future
What fascinates me most is how quickly things are evolving. A year ago, the copy-paste workflow was standard. Now I have an AI assistant that understands my project, can run commands, and integrates with external services.
The question is no longer if AI will change how we code - but how much and how fast.