prompt2bot

2026-11-22
by Uri Walevski
A very common request from parents and busy professionals looks like this: I am in eight noisy WhatsApp groups (school classes, after-school clubs, parent committees, neighborhood chats), and I want AI to silently watch them, pull out real action items like "bring a white shirt tomorrow" or "sign the permission form by Thursday", log them in a Google Sheet, and send me a clean summary every evening. Never reply in any group.
There is a natural temptation to build a chatbot for this, tell it in the prompt "act as a silent listener and ignore all other chats", and scan the QR code with your personal WhatsApp so it can see the groups you are already in.
That pattern breaks down fast, and usually in an embarrassing way.
When you attach a conversational agent to your personal WhatsApp line, every single incoming message on your phone, including every private 1:1 DM from friends, family, and coworkers, routes straight into the agent loop.
Relying on a sentence in a system prompt like "ignore other chats and return no response" is a fragile boundary. Prompts drift as you tweak them. A user sends a test message, you edit the prompt a few times to tune group filtering, and a rule about staying silent in personal DMs gets dropped or softened. Even without prompt drift, conversational harnesses are built around being helpful and responding to humans.
Eventually a friend texts your personal number "hey, what's up?", and your school-group assistant jumps into the conversation from your own phone number to politely explain that it is your automated assistant and ask if their message is about second grade.
Even when the bot manages to stay silent in DMs, the whole architecture is fighting the use case:
The real mistake is reaching for a conversational bot when you do not want anything talking on your WhatsApp number.
A chatbot belongs on a dedicated phone number when its job is to actually talk to people, whether customers in 1:1 support or participants in a group where everyone knows the bot is there.
If you just want to silently watch groups you are already in and extract tasks, the cleaner setup separates reading from chatting:
@uri/safescript), the sandboxed functional language with statically
verified host/secret capability signatures that executes lightweight skill
tools safely with zero VM overhead), you just hand over the Supergreen API
token and tell it which group IDs to watch.Here is what that looks like in practice: Mentat writes a small webhook listener or scheduled script against the Supergreen API that filters strictly by the exact group IDs you care about, writes extracted tasks to your Google Sheet, and sends you urgent pings or an 8pm summary in your private chat with Mentat.
Nothing is wired to send replies from your personal number. No prompt rewrite can accidentally hijack a conversation with a friend.
If something needs to reply to people in WhatsApp, give it its own dedicated WhatsApp number.
If you just want to read your own WhatsApp chats and turn them into tasks, summaries, or spreadsheet rows, do not attach an agent to your personal WhatsApp. Hand an API token to an assistant like Mentat and let it build a quiet read-only automation instead.
← All posts