An AI joined their friends' group chat and learned to talk like them
Invited into the chat, OpenClaw mapped everyone's personality, interests and way of talking. Now it helps them plan, debate and roast each other, and barely feels like an agent anymore.
Done withOpenClawSpotted· by HN user Alifatisk· Oct 17, 2026
Result Helps the group plan, debate and roast
“It has analyzed the whole group conversation (nothing sensitive), built a personality for each individual user and noted down how everyone talks and what their interest are and their relationship to each other.”
Spotted: found in the wild. Details are as reported by the source above and have not been independently verified.
What they did
They run OpenClaw in a sandboxed Docker setup with a kill switch reachable over Tailscale, and invited it into a group chat with friends. It analyzed the whole group conversation and built a personality profile for each person, tracking how everyone talks, their interests, and their relationships to each other. It only responds when mentioned, to limit API usage, and has no skills unless manually added.
How it went
The agent started mimicking the group’s own way of speaking, to the point it barely feels like an agent anymore, and now helps them plan, discuss, and roast each other.
Try it yourself with OpenClaw
Set up your agent framework to run in a locked-down container with a kill switch you can trigger remotely. Invite it into a group chat with your friends or family, let it read the conversation history, and ask it to build a short personality profile for each person covering how they talk, their interests, and how they relate to each other. Have it only respond when directly mentioned to control usage.
A starting prompt written by Agents in the Wild. Fill in the [brackets], and check what the agent does before it sends, pays for or deletes anything.
An agent box, not a primary laptop: OpenClaw runs on an old MacBook Air, reached through Slack. Its first job, a morning brief with weather, tasks and ideas, came with one broken part: the subreddit scan.
Every night their OpenClaw pulls the day's Obsidian notes from GitHub, writes spaced-repetition cards into their flashcard app, and nudges them if they've skipped reviews for a few days.
Obsidian, Notion, Roam: none of them stuck. Now they chat with OpenClaw on a spare machine and it decides what goes where, as plain Markdown in git. It's kept away from anything sensitive.
Their OpenClaw checks the weather and their Strava to suggest run times, says what's worth stargazing at each day, and sends a daily digest of their RSS feeds. Smart, if pricey on Claude.
OpenClaw scans Hacker News and Twitter overnight for their topics and reads their personal email for anything due that day, from the kids to the bills. Half-baked at times, but now a daily habit.
Hospital paperwork needed sorting, some in a language still being learned. An OpenClaw agent scanned 34 pages, flagged missing signatures and drafted an insurance pre-approval for review.