What can be said about bolt.new after 20 hours of use and building 2 apps:

• I managed to build a simple project to display data from ansible inventory files as a table. On the very first run it made a project that worked, showed everything needed, and did it nicely. It took 2–3 hours to figure out what I wanted and implement it. I would have spent 1–2 days on this myself.
• A mid-size project — a universal chatops bot based on LLM — didn’t work out, even though I have a detailed spec and I know how it should be structured. Bolt kept pulling toward a web app, even though that’s secondary. It tried to write in js instead of ts. Each iteration took a long time and was expensive. In the end I got tired of torturing it.
• Not suitable even for mid-size projects yet, because there are obvious issues with context size.
• Bolt is barely integrated with their own online IDE StackBlitz, to the point that code can be lost.
• The endless chat interface, in my opinion, is not the best idea, because the context length keeps growing.
• Sometimes you have to fight the model. For example, it kept trying to use gpt-4 instead of gpt-4o-mini, even though I explicitly wrote that. In general the model is very headstrong, and it’s hard to persuade it.
• It’s very tuned for code generation. When you ask how to do something, it suggests generating code.
• Generation speed is like modern models. You ask for a change and go do something. Sitting for 1, 2, 3 minutes watching generation isn’t very interesting.
• After each modification the app tries to run the app, but it’s unclear with which command. Eventually everything broke and after that I had to run the build by hand.
• It sees logs, but often doesn’t detect errors in them. You have to prompt it.
• You can quickly roll back changes made by the agent. Works great!
• It doesn’t generate tests unless you ask.
• Missing error highlighting right in code. But that’s again about IDE integration.
• It works much better with typescript than with javascript. Which makes sense, because it needs less context for that.
• It starts to get lost when there are many edits. At some point the static analyzer found 20 errors at once, and fixing them took about 10 generations.
• The basic paid account is enough for the whole day if the project is small.

This is the next level after copilot and even zed. It’s closest to how IDEs should look in the future — a full-fledged assistant that writes code from a prompt. But it’s still got childhood diseases. I’d like something faster, more integrated into the IDE, and more interactive. And preferably in my IDE locally, not in the cloud. It’s actually surprising that Jetbrains still hasn’t made anything like this and is busy with little things.