A developer known as Nate has built an AI system that automates his personal text messages, claiming a 95% success rate without needing edits, according to a post on his blog, noslopgrenade.com, discussed on Hacker News.
How the "Digital Ghost" Works
According to Nate's writeup, the system ingests the full history of a conversation from iMessage and Signal, builds a running summary for context, and generates replies in his voice using a local large language model — specifically a quantized version of Llama 3 70B (llama3:70b-instruct-q4_K_M). Running the model locally avoids sending personal conversations to cloud AI providers.
The Prompt Is Everything
Nate attributes the system's success to detailed prompt engineering that instructs the model to mimic his humor, tone, and emoji use. One instruction he includes, per his post, is to "never admit you're an AI," so replies read as genuinely his.
The core components, as described in the post:
- Local LLM:
llama3:70b-instruct-q4_K_M, run on his own hardware for privacy and cost control. - Conversation summary: a running summary maintained for long-term context.
- Persona prompt: detailed instructions matching his tone and style.
- Multi-platform integration: works across iMessage and Signal.
The Ethical Question
The project has drawn debate, including on Hacker News, over whether an AI standing in for a person in personal conversations is a legitimate way to stay in touch or a form of deception toward the people on the other end, who are not told they're talking to an AI. Nate has said he sees it as better than sending nothing at all when he's too busy to respond himself.
The tool is not a commercial product — it's a personal build shared on his blog — but it points to a broader trend of individuals running increasingly capable local LLMs to automate personal tasks previously thought to require a human touch.