We’ve talked previously about some of the various tasks that philosophers might have large language models help them with in their work, mainly focusing on the ethics of doing so. But what about the mechanics of doing so?

There are different ways to make use of AIs like ChatGPT. You could simply open it up and start chatting with it. That’s somewhat akin to talking with a stranger. That might be fine, but sometimes, perhaps especially when it comes to work, it is more useful to talk with someone who gets what you’re trying to do and knows how to help you with that, who understands how work in your area is supposed to proceed, who can prod you with incisive questions and give your constructive feedback. If that’s what you want, then you should consider customizing a GPT.

How do you do that? It’s not that hard, as Robert Bloomfield shows in the following guest post. In fact, he has created a “philosophical repertoire assistant” prototype to help you get started.

Dr. Bloomfield is a professor in the SC Johnson College of Business at Cornell University. His primary area of expertise is accounting, and he first developed the tool he’s writing about in this post—theory-driven GPTs—for work in accounting theory. He is, of course, sensitive to the differences between accounting and philosophy, but despite those differences, he shows how theory-driven GPTs could be of use to philosophers.

Best Next Moves: How Philosophers Can Make Good Use of Theory-Driven GPTs
by Robert Bloomfield

How should philosophers use GPTs? Daily Nous readers have seen many views on this, most recently from GPT5 itself responding to Justin’s discussion starter. I’m with Gus Skorburg on this: we can and should choose to use AI to learn. But Gus mostly focuses on better prompting. I am going to focus on GPT customization. OpenAI has designed ChatGPT so anyone can upload instructions and documents to direct the GPTs responses to all future prompts, and while some knowledge of coding is helpful, everything can be written in plain language. In this post I will describe how I customized a variety of theory-driven GPTs, including one for philosophers, each designed to help scholars find the “best next move” in their analysis. I will explain how they foster learning rather than undermining it, why they pose limited risk to academic integrity, and propose some best next moves for Daily Nous readers before they head into the comment section.

Machines and Moves

I will start by making three general points about customization.

1. Customization is an antidote to hype and moral panic. Hype and moral panic are both stoked by language that makes it seem like “AI” is one thing that can do anything. To paraphrase how Rodney Brooks puts it, that’s like saying “machines can do so much” and “machines will take all the jobs” without acknowledging that a machine that does a good job trimming a lawn does a lousy job trimming a beard. Customization makes clear that even GPTs (one narrow kind of AI) are not one thing—each GPT is its own thing, designed to be better at some tasks even though that makes worse at others. Uncustomized GPTs are designed to be better at what the typical user wants for their typical uses, which makes them worse at what a philosopher wants.
2. Customization works well for move-based analysis. Customization for scholarship works quite well when it is focused on supporting move-based analysis—breaking an issue down into small parts and handling them one at a time—because that is what GPTs already do, one token (lexical unit) at a time.
   • The first step in customization is thus to lay out a theory of “best next moves” given whatever theory you are working with—no technology involved.
   • The second step is to embed that theory in the GPT so that when it searches for the best next token to generate, it is doing so with clear guidance on what best next move that token is building.
3. Theory-driven customization keeps the scholar in the driver’s seat. When a scholar creates a theory-driven GPT they keep themselves in the driver’s seat. They aren’t just hoping the GPT will do what they want, they are articulating the theory they are using for analysis and evaluating the GPTs outputs in light of that theory. Not only does theory-driven customization make the GPT perform better, it also provides an audit trail. If the scholar did a good job articulating their theory, and the final output reflects a thorough and accurate theoretical analysis, that greatly reduces concerns the GPT might have made mistakes along the way or that the scholar used a machine to do their work for them: either the GPT was reliable or the scholar did a lot of careful checking.

Moves in Role-Centered Accounting Theory (RCAT)

Before turning to how philosophers can make use of theory-driven GPTs, I will share how I developed and made use of them in my area of expertise. The first step was to develop a theory, which I have done with Tamara Lambert (University of Manchester) and Marietta Peytcheva (Lehigh University). Role-Centered Accounting Theory (RCAT) provides an ontology that serves not just as a “language of business” but a “language of enterprise”, where an enterprise is any endeavor to achieve shared aspirations. (For example, scholarship is an enterprise endeavoring to ask better questions and generate better answers, among other aspirations.)

RCAT describes the challenges enterprises face, the tools they have to overcome them, and normative principles that describe the ways in which accounting can be better. RCAT analysis seeks to answer the countless and mostly noncontroversial empirical questions required to find a shared and justified understanding of a problem and its solution: Who has been assigned what obligations? Who benefits when those obligations are fulfilled? On what basis is a steward held accountable? Individual questions are pretty easy, but it takes a lot of answers to get to the solution, and it is important to follow the right sequence. We therefore designed a host of step-by-step protocols that lay out the best next move for different tasks. Practical analysis protocols move step-by-step toward identifying problems and their best solutions, while scholarly analysis protocols move step-by-step toward better problems to solve (for practitioners) and questions to ask (for scholars).

Theory-Driven Customization

We developed RCAT’s theory and protocols well before we thought about GPTs (we started in 2020). But when GPTs become so powerful and popular, the strong “best next move” connections inspired us to start customizing the RCAT Analysis Machine. In ChatGPT’s customization page (click the “Create” button on the top right of the main page), we pasted in just over a dozen very terse instructions requiring the GPT to “follow all instructions listed in [this part] of [this] translator file”, and uploaded three kinds of documents:

1. Theory Suite: a complete explanation of RCAT’s, its terms and how they relate to one another.
2. Protocol Suites: step-by-step guides for analysis.
3. Translator files: files that explain the details of each requirement in the main instructions.

The GPTs most important requirements in the translator files instruct the GPT to treat the Theory Suite and Protocol Suite as authoritative. Unlike a document uploaded as part of a prompt, documents uploaded in customization and deemed authoritative are always available for the GPT to access and the GPT gives their tokens highest priority for inclusion in its context window (the set of tokens included in processing). These documents are written in plain language so humans can understand the theory and follow the protocols on their own, but as authoritative documents they keep the GPT relentlessly focused on the best next move. Tokens relevant to that move (those from the last move, those leading to the next move) are given highest priority for processing and generation, while others are demoted until the state of analysis makes them relevant.

The translators also describe profiles that lay out the four jobs the GPT can switch among:

• Guide (asks protocol questions, doesn’t answer)
• Assistant (asks and tentatively answers)
• Questioner (asks tough questions to think about)
• Commenter (offers suggestions for improvement)

Three tools keep things stable and flexible:

• Anchors: the GPT summarizes progress at natural checkpoints and reuses those summaries so the thread doesn’t drift.
• Priorities: you can toggle faithfulness (stick tightly to RCAT), order (follow the protocol strictly), or fluency (brainstorm freely).
• Styles: you can direct the GPT to explain why it is doing what it is doing in RCAT terms (teacher), answer questions about RCAT (tutor), or highlight user options for improving GPT performance (power user)

Another requirement discourages the GPT from closing responses like “Do you want me to write that in a form suitable for a journal?”, instead encouraging responses that take people back to RCAT and analysis, like “You’ve raised an important distinction between personal and assigned assets. Do you want to clarify how that applies here or move on to the next step of the Cause Locator Protocol?”

These practices make the best next move more obvious to the GPT and keep the scholar in the driver’s seat.

Theory-Driven Customization for Philosophers

I have been pleased enough by the performance of the RCAT Analysis Machine that I decided to sketch a custom GPT to help with philosophy. Moves in accounting tend to be deliberative—we aren’t debating RCAT itself, just trying to find shared and justified understanding of each question it forces us to ask for a complete analysis. Moves in philosophy tend to be more focused on refining theory, and are thus more argumentative: stating a thesis, attacking it, defending it, and so on. Given that I am not a philosopher, I didn’t try to create my own theory of argumentative moves. Instead, I started with a GPT customized to be “the fastest way to learn anything hard” (the Universal Primer). I asked it to help me map common argumentative moves in philosophy, and to the eyes of this amateur philosopher it seemed good enough for a prototype. I then worked with the Universal Primer to convert the map into instructions to recognize and follow the repertoire of moves and toggle between four roles:

1. Teacher: teaches the repertoire.
2. Sparring Partner: deploys appropriate moves from the repertoire against your claims.
3. Tactician: helps you identify best next responses to a tough move.
4. Simulator: simulates an argument by alternating between opposing sides.

The whole process took about an hour. (It took far longer to write this post!) Here is the repertoire and customization code. You can see my first chat with it here. Maybe someone can use this to follow up on Justin’s 2017 post about Liam Kofi Bright’s call for higher standards in philosophical argumentation.

The Philosophical Repertoire Assistant is just a prototype. If you want to tailor this GPT to your own work, you will want to revise the instructions to improve the repertoire, and upload additional documents that lay out terminology and best next move protocols tailored to your needs, much like the Theory and Protocol Suites we used for the RCAT Analysis Machine.

Philosophers in applied fields should also consider re-customizing the RCAT Analysis Machine for their own needs. You can think of RCAT as a framework for working through the “easy” parts of applied ethics: if we know what we are trying to accomplish, RCAT analysis helps us figure out how. But RCAT offers no answers to what aspirations enterprises should have or how they should be prioritized, and its native claims speak only to the functions, circumstances, qualities, and normative principles of accounting practices. For a more complete analysis, the Machine includes a Theory Importer Protocol that helps scholars insert their own theoretical claims into the appropriate part of RCAT’s structure to allow an integrated analysis. We have already uploaded documents into the RCAT Analysis Machine that import theories of scholarship, data gathering, and GPT customization because we expect most of our target users to want those and we also feel qualified to articulate the theories. Philosophers would be far more qualified to import theories of applied ethics—if you want to do so, reach out and I’m glad to help.

One you get the hang of it, customization is easy and helpful. To prevent myself from being seduced by the flattery and agreeability built into OpenAI’s GPT, I’ve created Critic and Teacher, which draws from a range of frameworks for scholarly critique to toggle between relentless critic and patient teacher. You can read about it here. I also created a Music Finder GPT whose main job is to take a “seed” piece of music, analyze it across a dozen dimensions (genre, time period, arrangement, nationality, etc.) and offers suggestions of other pieces that vary in one dimension at a time either a lot (high heat) or a little (low heat). Its instructions are also good for less structured move-based searches. The other night my wife and I used Music Finder to introduce us to music from South Asia, starting with traditional folk music most familiar to American ears (little use of microtonality), which led us to some wonderful Bengali pentatonic folk music we had never even thought about listening to before.

Best next moves

No doubt the comment section will have plenty of argumentative moves, and hopefully some deliberative ones too. To prepare myself, I shared this document with the Philosophical Repertoire Assistant to ask what arguments to expect. As you can read here, it expects to see camps of Pragmatic Enthusiasts, Skeptical Critics, Integrity Worriers, Playful Experimenters, and Meta-Commentators (the last asking “what does this mean for philosophy as a discipline?”). But before you dive into commenting, consider these data-gathering moves:

1. Experiment. If you have a ChatGPT account (ideally Plus at $20/month), play around with the RCAT Analysis Machine, Philosophical Repertoire Assistant, Critic & Teacher, Universal Primer, or even Music Finder. A good first step is to ask any customized GPT “Why this GPT?” so you understand how to use it. To make sure you are seeing the influence of GPT customization it helps to turn off “access to saved memories”; otherwise, responses will also draw from your past chats. You can also check out sample prompts in the papers linked to on the front page of RCAT Central.
2. Read Chats. If you are unwilling or unable to experiment with GPTs, go to RCAT Central. The Crowdsourcing paper on the front page has not only sample prompts but also lots of shared chats toward the end. There are more shared chats in the Blog tab, including some on philosophy. (It turns out RCAT has deep connections to American pragmatism and ethics and offers some interesting applications of existentialist tensions.)
3. Read Customization Documents. The front page of RCAT Central has the full set of customization documents for the RCAT Analysis Machine. The instructions and translators are coded, but the others use plain language.

My coauthors and I have decided our own best next move is a crowdsourcing project: we are sharing with scholars RCAT, its protocols, the RCAT Analysis Machine, and its customization documents, and encouraging people to tell us how they use it and find it useful (or not). In the short time between submitting this guest post to Justin and getting his response, we received enough input to upgrade our GPT to the RCAT Entropy Managing Analysis Machine, which performs much better for reasons explained here. If you are willing to share experiences, successes, or failures, or want some guidance on customization, please reach out to me by email.