Build Your Own AI Research Assistant

Seminar concept note · by Sergueï Spetschinsky, PhD

The idea

At the heart of this seminar is a single idea: that a researcher can build their own AI-powered research assistant, a personal system, a kind of second brain, shaped to their own work and built in plain English, with no code. And that doing so has a real chance of making their work both more productive and more enjoyable.

Building that kind of system has a steep learning curve, especially when done in isolation and without clear conceptual and technical guidance. The goal of this seminar is to take the students through this first phase of setting up their infrastructure, before it starts proving its usefulness for their research work.

A semester-long research seminar

Unlike the shorter courses I teach, this one cannot be compressed into a few days. Building a research assistant takes time, practice, and a group to work with, so the seminar runs over an entire semester: an intensive in-person start, then regular online sessions in which participants show what they have built, bring the problems they have hit, and work through them together. In practice, it is a research seminar, run as a collaborative research group, and each participant demonstrates their finished assistant to the group at the end.

The era of agentic research

In the last year, AI products (like Claude Code and Codex) have crossed a technological threshold: from consumer-grade chatbots to powerful tools, which, if used with the right know-how, can serve as fully integrated systems capable of executing, at scale and great speed, most of the tasks a human could possibly do on a computer.

This technological revolution makes a genuinely AI-native research practice possible for the first time: agentic research, or the integration of complex AI infrastructure into the research process.

Agentic research is not just a new methodological option: it is a brand new field of research in its own right, which will take its own particular form and iteration in every single scientific field. So-called "soft" sciences are in fact, and contrary to what is often pictured by the researchers themselves, much better positioned to apply agentic AI methodology than so-called "hard" sciences. AI models are language-based, making them much more at ease with words than numbers, and that is why they are called LLMs, or large language models. Because of their linguistic, interpretive and qualitative nature, social sciences and humanities (Geisteswissenschaften in German, sciences humaines in French) are the ideal candidates for adopting agentic research processes.

With the right skills and the right system in place, a single researcher with a €200-a-month Claude subscription can now carry out, within days or weeks, scientific work that, not even two years ago, would have taken them months or years to do alone. For specific kinds of tasks at least, a lone PhD student can now work as if they were a lab director with an entire team of junior researchers under their orders.

This seminar is designed for social-science and humanities researchers who are highly motivated to explore that path.

An experimental and collaborative seminar

This is an experimental seminar by necessity. Not only is this a novel type of class, but we are spearheading a brand new, constantly evolving research field. AI tools, just like their application to agentic knowledge management systems, move extremely quickly, with fundamental shifts every few months.

For those reasons (but also because it is my own personal preference in teaching), it is designed as a hands-on, collaborative research group.

Participants from different disciplines are invited to bring their own projects, perspectives, research, expertise and ideas. This concerns not only their own field, but also the specific agentic system design they may bring with them or will gather along the way while building a system adapted to their own needs. They'll be required to do their own research, testing and experimentation, on which they will report during every class. Trials, errors and iterations are expected. And if I bring my own experience and systems, I assume that I'll learn as much from teaching the class and its participants as they'll learn from me. What makes this seminar worthwhile is the experimental knowledge we build together as much as the systems and competences that will be its outcome.

Outcomes

What each participant takes away, in working tools and in skills:

A working, self-designed, personalized AI research assistant, ready to be used in their research work and even in their private lives.
The ability to set up and run an AI-powered second brain over any kind of digitalized research material or knowledge base (notes, an in-progress dissertation, literature and text corpora, ethnographic or qualitative data).
A good grasp of the fundamental notions of agentic AI: system and meta-prompting, context management, agentic system building, sub-agent orchestration, the integration of external tools and data sources, automated workflows, and file-system and command-line (CLI) integration.
The skill to use and document AI responsibly, in line with their institution's guidance.
The competence to keep going: leaving the seminar set up to continue building on their own, in a position to integrate agentic research methodology ever further into their work; in other words, to become AI-native researchers in their own right.

Not a set of facts about AI, then, but a working system and the skills to keep developing it.

What is a personalized AI research assistant?

This chapter describes the kind of system the seminar helps each participant build: the same kind of system I myself work in every day.

At its core, an AI research assistant is a personal knowledge base: a collection of notes written in plain language, kept as Markdown files (a simple, durable text format) in folders, each note carrying a small block of structured information at its top, a kind of customizable metadata called frontmatter, written in YAML, a writing format that is easy to read both for humans and LLMs. This is the idea often called a "second brain" or personal knowledge management. The whole base can be displayed and navigated easily and elegantly through a free program called Obsidian, which turns the folders of notes into a browsable, visually linked map of one's knowledge. On top of it sits Claude Code, the agentic AI tool through which the knowledge base is built and run. The rules that turn the knowledge base into an agentic system, a semi-autonomous research assistant, are also written in plain English. Taken together, the knowledge base plus the AI working over it amounts to a kind of personal AI operating system: a single environment, shaped to one researcher, in which one's materials and one's AI tools live and work together. It can hold any knowledge that can be put into words: literature, reading notes, ideas, drafts, data, and records of one's own thinking. And a well-kept base is fully legible to the AI working within it: through that work, it becomes, in effect, a prosthetic brain.

Some of what it makes possible, depending on the capabilities one decides to build:

Remembering and connecting. The base holds everything that has gone into it, kept live and within reach, an extension of our own famously limited working memory. Not only can it dig up a single line or quote from a huge mass of material on nothing more than a vague command, it can also surface adjacent or connected information on its own, unprompted, in the middle of writing or thinking an idea through.
Capturing ideas on the go. Information goes in from anywhere, by voice, in seconds. Out doing the groceries, an idea about a chapter strikes; one speaks it into the phone, however messily, and the assistant cleans it up, connects it to related ideas, and files it where it belongs.
Ingesting research material automatically. Whatever feeds the research becomes clean notes in the base: PDFs, scanned documents, web pages, audio and video recordings. Drop in a folder of hundreds of papers and it comes back converted, every single paper summarised, tagged, organised, given full frontmatter metadata, and connected, ready to be queried or drawn on as context.
Finding patterns. Spotting patterns across a large body of material is one of the things these models do best, and for qualitative researchers it is a powerful instrument. Across hundreds of interview transcripts it surfaces the themes that keep recurring, or traces how a single notion shifts as it moves through a body of documents.
Producing new knowledge by connection. The assistant draws together what is already there, the curated knowledge in one's own base, to help produce new writing. From a few spoken commands it takes a chapter draft, sets it against a large literature corpus to see what bears on it, and pulls in the relevant findings from one's own field data.
Writing, synthesising and presenting. Outputs are produced to one's own standards: a defined writing style, format, structure, and templates, applied the same way every time. A scattered set of reading notes comes back as a structured literature review, in one's own template and style, ready to revise.
Reviewing, feedback and critique. The same standards then work in reverse: the assistant reviews a draft against the precise criteria one has set and gives feedback in those terms, and as a thinking partner it pushes back. Asked to argue the opposite case, it exposes the weak step in a chapter's reasoning before a reviewer does.
Reaching out to the open web. The assistant is not sealed inside its own notes. It runs deep web searches, querying hundreds of sites in minutes and summarising what they return.
Integrating the research toolset. Through the command line and through MCP (an emerging open standard for plugging an assistant into outside software), it connects to a fast-growing range of tools and operates them for you: reference managers like Zotero; the major bibliographic databases, hundreds of millions of works searchable without leaving the base (OpenAlex, Crossref, Semantic Scholar); spreadsheets and statistical environments (Excel and Google Sheets, R and Python); survey platforms such as Qualtrics; and everyday tools like email.

What people build it into. Most people start from the same place, a note and idea system, and grow it into something shaped to their own project:

A note and idea system (where almost everyone starts). A place to capture notes and ideas and accumulate one's own knowledge over time. It is the foundation, and further layers are added on top of it.
A writing system for the dissertation. The note system extended into the writing itself: chapters, arguments, sources, and notes held together so the assistant is aware of the whole as one writes.
A literature store. A body of reading, summarised and indexed, so the assistant keeps a working view of the field.
Ethnographic or field material. A disciplined way to automatically encode and query a very large body of interviews, observations, and other field material.
A corpus under study. For work whose object is itself a body of texts, the whole corpus can be embedded, represented so the assistant can search it by meaning rather than by keyword alone.
A mixed-methods project. Where the research joins qualitative and quantitative work, the assistant holds both sides in one place, so the two can be read against each other.
A mix of these. In practice most systems combine several: a dissertation at the centre, a literature store around it, and field data feeding both.

The system is built to last beyond any one tool. We build it in Claude Code now, but it is kept as independent as possible from that particular tool and company: the knowledge stays as plain text the researcher owns, and the setup can be ported to other AI providers down the line, and, as local models improve, to models that run privately on one's own computer.