If you live outside the city, you absolutely have to vibe code. Especially in winter.
I got myself a Claude Code subscription a while ago, and it works pretty great. Describe the task, set the necessary permissions so it works without unnecessary questions, start it. And that's it, you can go outside and shovel snow. If there's a lot of snow, you can set several tasks at once.
Lately I've been using Claude for all it's worth. Judging by the regular messages about exceeding the daily and weekly limits, it can't get any more intensive.
To make the limits last a little bit longer, I use this simple method.
- First I create a CLAUDE.md file for the project, where I describe information about the project that's useful for the little AI. What technologies, practices, tests, how to clear the cache and where to find the right containers. The neural network takes this file into account at the start of every new session. In principle, you can also ask Claude to write this file with the /init command, but it doesn't understand everything about the project right off the bat.
MD stands for Markdown, a convenient way to format text in plain text files. Convenient for both robots and humans.
- Then I write separate md files with tasks. I describe the goal of the task, the details, additional info.
- Then I switch Claude into planning mode and say: read the task file and draw up a detailed plan for implementing it.
- I read the resulting plan, click over to auto-edit mode, start it and go shovel snow.
- Some people recommend explicitly asking the little AI to record the progress of the plan in a separate file, so you can come back to it in case of a failure. But as far as I've noticed, it writes everything to files under the hood anyway, and there are no problems continuing an interrupted session.
- When Claude is done and the snow is done, everything it wrote needs to be checked by hand, debugged and committed. I'll write about that part next time.
As I mentioned above, the daily and weekly limits run out faster than I'd like. To save a few tokens and stretch carefree vibe coding just a little, I recommend looking through the session log yourself after two or three completed tasks, or even asking the agent to reflect. Namely: to update the contents of CLAUDE.md based on the logs. The agent often fumbles at first and doesn't immediately understand exactly which command is the right one for running tests, or that it doesn't need to go into a container to build the frontend.
These are the commands you need to correct, writing down in CLAUDE.md the variant the agent managed to complete the task with. So it doesn't go through the options all over again in every new session.
I've used both the console claude and the VS Code plugin. The console one is cooler, with more features and all sorts of commands. Besides, you can run it with checks turned off, so it doesn't ask whether it may run rm -Rf * and just does it. But the plugin is fine too. The little AI works conscientiously in both, and the monthly subscription, I'd say, paid for itself in the first 3 hours of use.
So, happy snowy vibe coding! And in summer we'll get the lawnmower out and do some vibe debugging.