Groundwork

A knowledge graph and a Claude agent with full read/write access. Built it for myself.

Two tables: nodes and edges. Properties are freeform JSON. No rigid schemas, no app-specific tables.

Everything is a node. Beliefs, people, feed items, signals, projects, the page you're reading now.

The agent sits on top of the graph. Skills are markdown files. One for ingesting a profile, one for surfacing what I read, one for drafting content, one for working through a thesis. A few hundred lines each. The agent picks which one to use.

What I use it for

• Brainstorming. I dump something I'm chewing on. The agent expands it, finds what already connects in the graph, and lets me follow threads.

  

• Loading my own data. What matters to me ends up as a node. Beliefs, values, what energises me, what drains me, people I want to keep close, things I've read.

• A personalised feed. Hacker News, RSS, a handful of other sources. Pulled in, scored against the graph, ranked. What I scroll on my phone is shaped by what I actually care about.

  

• Writing and the website. Every page on qni.dk, including this one, is a node. I draft, edit and publish through the agent.

  

• Background agents. Cron-scheduled runs that work while I'm away. Pulling the feed. Ranking it. Cleaning out what's gone stale. Pre-drafting things I'll come back to.

  

• Idea validation. When I want to dig into a specific idea, I have skills for research and outreach. One thing the graph is useful for, not the point of it.

What is public, what is not

What you see on qni.dk is a few dozen nodes. This page, the blog, projects, notes. The graph itself is around 600 nodes and 600 edges. Drafts, raw research, journal entries, signals from things I read, contacts, transcripts. They sit in the same database, connected to the same nodes. Publishing is flipping a status.

Still early. The architecture is deliberately boring. Postgres on Neon, a Hono worker on Cloudflare, the Claude Agent SDK, MCP tools for graph access.