What this page is

This page shows one example of how a thesis can be run with an AI agent that has access to a Notion workspace. It lays out the structure and architecture of the setup: a reference catalog, an empirical data tracker, and a short section written for agents.

The idea is a division of labor. The person reads, judges and decides what an idea means for the argument, while the agent files, searches, formats and checks. Treat it as a starting point and a worked example: the topic, chapters, fields and habits here are one person's choices, and anyone building their own process will end up with different ones. The thesis this setup supported is available at urn.fi/URN:NBN:fi-fe20260923128190.

<aside> 🧪

Example data only. The rows below come from a real working setup but were trimmed and generalized, and the interview rows were anonymized. Statuses and labels were adjusted so that every option appears somewhere, and nothing is connected to the original databases.

</aside>

Part 1: reference catalog

The catalog turns reading into citable material. It has two linked databases: one row per source, and one row per idea taken from a source.

Every claim in the draft can be traced back to a source, a page and a reason for using it. Because the human reasoning is stored next to the evidence, the agent never has to guess.

Structure and principles

Sources hold what a source is, and insights hold why it matters. Each insight points to one source and each source lists its insights, while a chapter, a stance (for example supporting, contradicting or methodological) and a strength on each insight let the agent order material the way a writer needs it.

A few principles do most of the work:

• Insights come from the person's own reading. The agent writes only the neutral summary of a source. The annotation, which holds the reasoning, belongs to the person.
• Unread is flagged in the data. If only the agent has read a source, its summary starts with PLACEHOLDER and a warning not to cite it. A source with zero insights is not a defect, since it often means unread or background only.
• Disagreement and weak sources stay in. Contradicting insights are kept, and sources that are not citable are labelled so they cannot slip into the reference list.
• Usage is tracked. A Used in thesis field records where an idea appears in the draft. A filtered view of the sources where it is filled, sorted by authors, works as the working reference list.

Example data: sources

The table below holds 15 example sources of many types and reading statuses. The tabs above the table show the same rows as a working reference list, as a reading board and as a list of sources that have no insights yet.

Sources (example data)

Example data: insights

The table below holds 17 example insights, each linked to one of the sources above. Open any row to read the annotation, and use the tabs to see the insights by chapter, by stance, or as strong ideas that are not used yet.