I built aipatch after running into repeated problems with AI coding tools that would eventually fail to apply patches correctly or ignore parts of the instructions. They often kept doing things I didn’t ask for, especially when the context got even slightly unusual.

I wanted something simpler and more explicit: a tool where I control exactly what goes into the LLM prompt, and the model responds with deterministic SEARCH/REPLACE blocks that can be applied automatically.

The main idea behind aipatch is: - You choose the context manually (from one or multiple projects). - You send it to any LLM you prefer. - The model outputs structured patch blocks. - aipatch applies them to the codebase and logs everything.

The part that turned out most useful for me is multi-project prompting. I often need to update backend + frontend + documentation together, or compare two git commits, or port a feature from a small prototype into a larger repo. Existing tools don’t handle this well, so aipatch lets you give each project its own ID and merge them into a single prompt.

I also found that the best “prompt” is often just another working project that already behaves the way you want your project to behave. Including a small prototype or reference repo in the context gives the LLM a clear example and produces much more accurate patches than trying to describe everything in plain text.

A short demo video ("I improved aipatch using aipatch") is here: https://youtu.be/xho0pMKPu14

GitHub: https://github.com/axife/aipatch

It’s a CLI tool written in Python, works with any LLM, and doesn’t require an editor integration or an account. I’d appreciate feedback from anyone who works with LLM-assisted development or multi-repo workflows.