studyhall 1.2 - Notion as an AI Knowledge Harness / Skills Collector / Notion AI Workers

Link to Folder with Everything

2026.10.07.studyhall.1.2

Recording of the Video

Watch on Google Drive

Build It Yourself — The Guide

Everything from the session is public: the eight copy-paste prompts, both versions of the Notion Collector skill — SKILL.md (single-file) and notion-collector.zip (multi-file) — and the full walkthrough.

Slides

21 slides: the observer flow, MCP in one breath, Notion for people who have never opened it, and the eight-step build.

Open the deck

Summary

This one was hands-on: by the end, everyone had a working knowledge harness in their own Notion workspace, driven from Claude. Rahul ran the build; Jabari Grubb — Notion Ambassador, with early access to the product roadmap — demoed what Notion's own AI can do without any external agent at all.

The thesis

The bottleneck in AI workflows is no longer execution — it's knowledge structure. An agent without a structured knowledge base produces generic or wrong results, no matter how good the model is. So the session's real subject was where an agent should keep what it learns, and why the answer for mixed human-agent teams is a database with status, relations, and views — not a folder of files. Git and Obsidian are excellent stores; what Notion adds is a work queue agents can fill and non-technical people can review, built in.

The build: a skills collector in eight steps

Using the tekt.observer verbs — Explore, Seek, Organize, Understand — the group built a harness in Claude that researches places publishing agent skills, collects them into a Notion database, tags and summarizes them, and then stops. The stop is the point: a human review gate sits between Organize and Understand, because a collected skill is instructions written by a stranger. Hidden prompt-injection payloads inside skill files can hijack an agent, so nothing gets used until a person has read it and dragged the card to Approved. Each stage reads one status and writes the next, which means any stage can be rerun, handed to a different agent, or done by hand — and next week's run needs one prompt and no chat history, because the databases carry the memory.

Notion's native AI workers

Jabari's demo covered the other direction: agents that live inside Notion (Business plan, ~$20/month). Workers take imported skills via a "Use with AI" button, ship with explicit governance — scope, permissions, run schedule, guardrails shown before anything executes — and log every change with lineage, so "when did this break?" stops being a mystery. He had a worker research and populate a knowledge table live, with every modified page listed inline afterward.

Tricks worth stealing

Obi's workspace switch: name two Notion MCP servers identically in Claude Code to toggle between personal and work workspaces. Rahul's image pipeline: Flux.ai as an MCP server inside Claude, flagged as a serious upgrade for generating visuals mid-workflow. And the quiet kicker — intelcraft.ai itself is Notion as a CMS: these very posts are Notion pages published as a live site.

Next steps

• [Everyone] Download the Notion Collector skill from the session page (ZIP + prompts) and run the loop on your own topic — skills, links, repos, MCP servers, papers, anything with a URL.
• [Everyone] Bring what you built to a future office hours — use cases wanted.
• [Rahul Singh] Skills zip and instructions shared with participants via the session page.
• [Jabari Grubb] Keep the group posted on upcoming Notion releases and sessions.
• Worth a trial: Notion Business plan to test native AI workers and their governance controls.

Details

• Session: studyhall 1.2, October 7, 2026. Rahul Singh and Jabari Grubb presenting. Recorded on Meet, with Granola and Fireflies as backup scribes.
• Format: fifteen minutes of context (observer flow, MCP, Notion fundamentals, the competitive landscape), then 45+ minutes of hands-on build in Claude.
• The schema: two databases — Sources (the allow-list: Proposed, Active, Paused, Dead) and Skills (the inbox and library: New, Organized, Approved, Rejected) — plus a Pipeline board grouped by status. The agent only ever writes Proposed; a person promotes to Active. That is why the agent cannot wander.
• Notion vs the field: compared against Trello, Airtable, Confluence, and Jira — Notion's edge is combining wikis, project management, databases, automation, and publishing in one no-code platform. Caveats on the table: no offline support on the roadmap, and native automation is thinner than Zapier or Make.
• Where it's heading: agent-to-agent communication beyond MCP is emerging, and a governed, structured hub like Notion fits that model — the session's analogy was a message board where agents collaborate on a shared thread, with real permissions.
• This summary was written by Claude from the Granola notes and the Gemini transcript (both in the Drive folder).