Every developer's Twitter feed is a firehose of new AI tools promising to revolutionize coding. I've tried dozens, but only a handful have earned a permanent spot in my workflow. The AI tools I actually use as a developer are the ones that solve specific, repetitive problems without breaking my concentration or requiring constant context-switching. The rest were distractions I eventually dropped.
How I use GitHub Copilot for real work
I don't use Copilot as a crutch to write entire features. I use it as a supercharged autocomplete for boilerplate and tedious patterns. It excels at generating repetitive code I already know how to write but don't want to type.
For instance, when I need a new API route handler in a Next.js project, I'll start typing the function signature and let Copilot fill in the standard structure. It saves me from writing the same try-catch blocks and validation skeletons repeatedly.
// I type: export async function POST(request: Request) {
// Copilot suggests the rest:
export async function POST(request: Request) {
try {
const body = await request.json();
// Validate input
if (!body.email) {
return new Response('Email is required', { status: 400 });
}
// Process data...
return new Response(JSON.stringify({ success: true }), { status: 200 });
} catch (error) {
console.error('POST error:', error);
return new Response('Internal server error', { status: 500 });
}
}
The key is treating its suggestions as a first draft. I always review and edit. It’s a productivity multiplier, not a replacement for understanding the code.
Why I switched from ChatGPT to Cursor IDE
I used ChatGPT heavily for brainstorming and debugging, but constantly switching between my editor and a browser tab destroyed my flow. Cursor IDE changed that by embedding an AI agent directly into my editor.
Now, when I encounter a confusing error or need to refactor a component, I can select the code, hit Cmd+K, and ask questions in context. The agent has access to my project files, so I can ask things like, "How does the authentication logic in lib/auth.ts interact with this page?" It provides answers based on my actual codebase, not generic examples.
This context-awareness is the killer feature. It turns the AI from a general-purpose chatbot into a specialized teammate who knows the project. I dropped standalone ChatGPT for coding tasks because the friction of copying code and explaining context was too high.
Do you really need an AI code reviewer?
I tried several AI-powered code review tools that promised to catch bugs and suggest improvements. Most were underwhelming. They generated generic feedback about adding comments or following conventions I'd already consciously broken for a reason.
The noise-to-signal ratio was too high. I found myself dismissing obvious suggestions, which trained me to ignore the tool entirely. For now, I've dropped dedicated AI reviewers.
Instead, I use Cursor's agent for targeted reviews. Before opening a pull request, I'll ask it to "review this hook for potential bugs or performance issues." Because it understands the surrounding code, its feedback is more relevant. It’s an on-demand consultant, not a noisy gatekeeper.
The AI tool I use for writing, not coding
Most of my dropped tools were marketing hype, but one non-coding tool stuck: Claude for writing. As a founder, I need to write clear documentation, blog posts (like the ones on suhailroushan.com), and product descriptions.
Claude excels at restructuring my technical drafts into clearer, more concise prose. I write a rough first pass with all the technical details, then ask Claude to "make this more readable for a junior developer" or "shorten this by 30% without losing key points." It’s an exceptional editor that helps me communicate ideas without getting bogged down in wording.
I don't use it to generate ideas from scratch. The input must be my own thoughts and structure. The AI then polishes the language, which is a task I'm happy to delegate.
When should you drop an AI tool?
You should drop an AI tool when it creates more work than it saves. If you're spending more time correcting its output, managing its context, or filtering its suggestions than you would have spent doing the task manually, the tool has failed.
I dropped tools that required extensive configuration, produced inconsistent results, or demanded that I adapt my workflow to their limitations. The best tools integrate seamlessly and solve a problem so well that you forget they're AI.
My workflow now is lean: Copilot for boilerplate, Cursor for in-context problem-solving, and Claude for polishing writing. Everything else was a temporary experiment. The real value isn't in trying every new tool; it's in finding the few that genuinely reduce friction and letting the rest go.
Find the one or two tools that fit into your existing workflow so naturally you forget they're there.