diff --git a/README.md b/README.md
index 837a6112d..06a8b1be5 100644
--- a/README.md
+++ b/README.md
@@ -41,10 +41,28 @@ Comment.find_each(&:refresh_search_embedding)
 
 ### Tests
 
+#### AI Requests
+
 For testing OpenAI API requests, we use [VCR](https://github.com/vcr/vcr). If you want to test AI features exercising the API, you need to place the `config/credentials/test.key`
 that you can get from 1Password in "Fizzy - test.key". Then, when running tests that use Open AI API, you must either set the env variable `VCR_RECORD=1`
 or to add `vcr_record!` to the test. See `VcrTestHelper`. Due to Open AI quotas, you may need to limit the parallelization. E.g: `PARALLEL_WORKERS=2`
 
+You can regenerate all the VCR fixtures with:
+
+```bash
+VCR_RECORD=1 PARALLEL_WORKERS=2 bin/rails test
+```
+
+A typical scenario is making modifications LLM prompts. You need to:
+
+1. Make the prompt changes.
+2. Run the test with `VCR_RECORD=1` set.
+3. Make sure that the test passes by running the test normally (without setting VCR_RECORD=1).
+4. Commit the YML fixtures that VCR has generated. That way, the next time you run the test, it will run fast without performing any request.
+
+Notice that if you pass changing data to the prompt, this mechanism won't work. E.g: if you pass data timestamps.
+You need to make sure those timestamps are always the same across executions.
+
 ## Environments
 
 Fizzy is deployed with Kamal. You'll need to have the 1Password CLI set up in order to access the secrets that are used when deploying. Provided you have that, it should be as simple as `bin/kamal deploy` to the correct environment.