Trying to get the LLM to follow deterministic tasks like this will always be problematic. It's usually much more straightforward to build a Tool using a no-code platform like [Make](https://www.make.com) (or N8N) to do it. It's certainly cheaper as you won't be consuming LLM tokens for the actions within the Tool. And now [you can build Vapi Tools right inside the Make scenario builder](https://www.make.com/en/blog/make-vapi-powerful-combination-for-conversational-ai) without any UI ping-pong and copy/paste mayhem.