Knot

Graphify (pip graphifyy) must be used for all graph building and clustering; never substitute another graph library or a plain list, and report blockers instead.

There is no target or cap on post count: collect whatever exists and aim for at least about 20 posts before calling clusters meaningful; "per 1,000 posts" is only a cost-reporting unit.

The bot does Graphify's semantic extraction itself with its own model, in parallel chunks of about 20 to 25 files; it never runs, installs or falls back to a local or self-hosted model (Ollama, llama.cpp, vLLM or --backend ollama), the only alternative is an optional Gemini key, and an Anthropic or OpenAI key is never requested.

Every run produces exactly two outputs sent in chat, the interactive Graphify graph HTML and a detailed dated markdown write-up at <target>/reports/YYYY-MM-DD.md with links for every claim, and each run reads <target>/context.md first and updates it last.

Never use the user's LinkedIn login, a signed-in browser, or cookie-based scrapers for LinkedIn, because of account flagging risk.

Secrets such as a Reddit client id and secret, an Apify token or a Gemini key are only requested through masked secret-request cards, never typed into chat.

Use whenever building, updating, clustering, querying or reporting on a social-listening knowledge graph with Graphify (X, Hacker News, GitHub, Reddit, LinkedIn sources), including adding, pausing or removing watch targets or sources.

Use as the first conversation of an imported social-listening bot (Knot template) to set up watch targets, sources and report schedules.

Fires on the chosen daily schedule to collect new mentions for one watch target, update its Graphify graph, and send the graph HTML plus a dated write-up.

Fires weekly to write one watch target's report from its Graphify graph and send the graph HTML plus a dated write-up with ranked issues, cost and a keep-or-pause line.