Terraform Community Scout

This example applies the monitor-on-a-cycle pattern to social listening: an agent that watches public conversations for places where your product genuinely helps, drafts a useful contribution, and suggests it to the team on Slack for a human to post. It is built on ChatBotKit's first-class Reddit abilities for reading (search and feeds work over RSS with no authentication) and the Slack integration's start-conversation ability for the hand-off.

The boundaries stay clean. Finding threads and reading them are tool calls; a small stdlib script dedups discovered threads so the same conversation is never re-scored or re-suggested; and the judgment - whether a thread is worth engaging, and what a genuinely helpful reply says - lives in SKILL.md playbooks the agent reads and follows. A watchlist.md config holds the product, keywords, subreddits, the team/Slack routing, and the engagement rules.

The gate is structural. Because this is an automated agent, the human-in-the-loop is enforced by the toolset rather than a prompt: the scout has read-only Reddit tools and a Slack-suggest tool, and no Reddit-post tool at all. It physically cannot reply on Reddit - it can only suggest to the team, and a human posts. This matters more here than almost anywhere, because auto-posting marketing to Reddit is how accounts get banned and brands get burned.

The whole thing is built to be welcome rather than spammy - a high relevance bar, one tailored reply per thread, affiliation disclosed every time, and subreddit rules respected. Other sources (LinkedIn, a news API, or X via the X API) can feed the same mention pipeline without changing the rest of it.