Im getting so hyped about setting up Claude Code to handle my unit testing workflow but honestly Im hitting a bit of a wall with the implementation details. Ive been trying to offload my boring test suite maintenance for my current freelance project—its a pretty standard Next.js app with a ton of backend logic—and I keep seeing people on X and Discord say that its all about writing perfect system prompts or just letting the agent run wild in the repo. One guy swore by creating custom tool definitions but then another thread said that just messes up the context window and makes the AI hallucinates tests that dont actually run. I feel like those answers are way too high-level and dont really touch on how to keep the tests from breaking every time I change a helper function.

Here is what I need to pull off:

• I need it to handle Vitest since thats what my codebase is built on.
• It has to be able to read existing mocks because I have so many custom ones that I dont want to rewrite.
• Ideally, it should run the tests in watch mode and actually fix them when they fail instead of just telling me what went wrong.
• My budget is basically just the Claude API costs, so I cant afford any fancy enterprise agents that cost hundreds a month.

Im based in Chicago and I have a hard deadline to get this feature release out by next Friday so I really dont have time to spend days debugging the agent itself. Has anyone actually figured out the specific skills or best practices to stop it from going in circles when it hits a failing test? Like, do you give it read-only access to the source and just let it write to the test files, or is there a specific way to prompt it so it actually understands the dependency injection I have set up? Just feels like every time I try it just writes a bunch of junk tests that dont even import my modules right...