67% of developers are using AI IDEs wrong. Here's how to effectively code using AI and AI IDEs for free.
This workflow helped me save hours and ship faster.
My Setup
I use Windsurf as my IDE since their SWE-1 is free for now, and sometimes I take my code to Claude as it's a smarter model. Cursor, Cline, Claude Code are all good options.
Planning the Project
Before writing any code, I mull over the architecture with respect to inputs, APIs, database setup, etc.
I often pose this as a discussion with ChatGPT and go back and forth until the architecture and flow make sense, then ask it to generate a prompt/PRD for my IDE. You can do this with any LLM—the output won't be drastically different.
Moving to the IDE
I paste the prompt in Windsurf, and here's where most folks mess up: instead of asking it to directly write code, I ask it to give me its plan of implementation.
These models still have a relatively small context length (~150k–200k tokens) so using it effectively is key. When SWE-1 gives me its plan, I can spot and fix problems early and save tons of time debugging AI slop. This does require some SWE experience (specifying API schemas, DB setup, etc).
Building the App
If it's a small project/MVP, you can one-shot the implementation. But I think it's better to break it down: build smaller components, review and test.
If the code works, congratulations! Now COMMIT and PUSH to git. For the love of god, DON'T MISS THIS STEP.
If it doesn't, DON'T just say "debug this." Read the terminal logs, understand the error, and give clear instructions (file name, function, line, etc). This prevents the IDE from going on a wild chase and burning tokens. Remember, preserve context length as much as possible.
Throughout this process you'll have to read docs, since they're constantly updated and models can't keep up. Web search/crawl MCPs may help, but I've had mixed results so I just do it myself.
The Bottom Line
Repeat this with human oversight and you'll have an app that works and a codebase that isn't complete slop.
Honestly, this approach is simple and straightforward when coding with AI. If you're using Claude Code, there are even more creative ways with their sub-agents, but I haven't tried them yet.
AI isn't replacing SWEs today, and I wouldn't trust it for end-to-end builds—but it's probably the greatest tool an engineer can have.
P.S. This was basically a brain dump to myself :)