Separate an assistant's action request from its WhatsApp reply
At a glance
Satvik Srivastava's Andro walkthrough separates Gemini's conversational text from structured actions before routing either output. This workflow is listed under Assistant Development, Structured Actions, Message Routing.
- Agents / tools
- Python, Gemini API, WhatsApp, Andro
SourceBuilding a tool for these workflows? See weekly rates.
What they did
1) Have a Python assistant with a Gemini client, a WhatsApp sender, and handlers for the actions it supports. Import re and json for response parsing. 2) Document each supported action and its JSON shape in the system instructions. Have Gemini put an action object inside <action> and </action> after its conversational reply. 3) Search the response with re.compile(r'<action>\s*(\{.*?})\s*</action>', re.DOTALL). If there is no match, return the original text and no action. Otherwise, strip the text before the match and decode match.group(1) with json.loads. 4) Send only the cleaned text to WhatsApp and pass the decoded dictionary to the appropriate action handler.
Worth knowing
This is one component of Andro, not a complete reminder service. The author's WhatsApp integration uses unofficial browser automation, and Gemini calls leave the local machine.