The Dojo

Can AI-based simulators help junior examiners speed run years of on-the-job training?

There is a scene in The Matrix where the protagonist, played by Keanu Reeves, needs to learn how to fight. A wire is plugged into the back of his head and a computer program which simulates martial arts training is uploaded to his brain. When it’s done he opens his eyes and says, “I know kung-fu.”

Accelerated simulated learning seems like a sci-fi dream, but in the age of AI it may be a necessity for junior knowledge workers. As AI tools take over more and more of the labor-intensive grunt work across enterprises, low stakes opportunities to learn on the job will shrink. This has prompted some to begin asking, “How are we going to equip and train the next generation of workers and leaders?”

My strong prior is that standard classroom teaching, certification programs, and self-study modules are not going to be up to the task. Too much of the development of critical thinking and judgment — which require good sensemaking and intuition — is acquired on the job through interactions and observations, not knowledge acquisition and study alone.

In the past, those interactions and observations were driven by external forces — eg., meeting and reporting cadences — effectively dictating the pace at which staff could acquire meaningful work experience.

What if a simulator could speed that up? What if we could create the knowledge work version of the kung-fu simulator that speed trained Neo in the Matrix?

The build

This experiment is based, in part, on my own experience.

After a couple of years at the SEC, I led a value-at-risk horizontal exam across what were then the major investment banks (Bear Stearns, Goldman Sachs, Lehman Brothers, Merrill Lynch, and Morgan Stanley). While I had studied derivatives pricing, the Greeks, risk management, and VaR in graduate school, none of those textbooks and classes could compete with what I was able to learn over several years of engaging with market risk managers, going through real life risk reports, reviewing actual VaR model documentation, and studying under an experienced, supportive boss with highly talented colleagues. A simulator needs to capture that.

I used Claude Opus 4.6 and GPT 5.4 in parallel to plan out the building of the VaR examination simulator. The initial prompt reads:

As AI adoption expands at enterprises, including regulatory agencies, junior employees are likely to have fewer natural opportunities to learn on the job. Using a coding agent like Claude Code or Codex, I would like to build a "simulator" that condenses those previously natural opportunities, which may have played out over years, into a series of curated experiences in a synthetic environment that feels realistic and enables junior employees to learn and absorb the same amount but in a much shorter period of time. (This would be the knowledge work version of the kung-fu training simulator from the first Matrix movie when Neo fights Morpheus.). I would like to start with value-at-risk -- i.e., a topic focused simulator that a junior employee could inhabit so that by the end, the employee would be able to effectively lead a VaR exam soup to nuts at a GSIB. Draft a plan for developing such a simulator. Don't make anything yet. I will iterate with you on the plan and then ask you to produce the precise instruction for Claude Code to build a prototype.

Note, I didn’t specify much. There are no references to rules or exam manuals or handbooks (just a parenthetical reference to The Matrix!). Rather, I lay out the overarching objective, signal that it needs to be built by a coding agent, and request “a plan for developing such a simulator”. I’ve found that this approach — of doing the planning via the chatbot interface and doing the building through the coding agent — works well.

Claude and ChatGPT diverged in interesting ways.

Claude’s initial response included a section up front titled “Design Philosophy.” It is worth reading in full:

Claude suggested building something centered on a narrative with increasingly challenging steps, complete with live dynamic conversations with a market risk manager, a suite of synthetic documents, and adaptive feedback from a boss. The inclusion of NPCs (non-player characters) was not something I was expecting, but is a key part of the simulation.

The full back and forth of my planning session with Claude Opus 4.6 is here.

With the same initial prompt, GPT 5.4 built a simulator that tracks more closely to examination handbooks and is oriented to helping users reach mastery across each component of the exam process. As a result, the supporting planning docs are much more expansive and detailed, and the narrative plays more of a supporting role (though NPCs are also used). The full planning conversation with GPT 5.4 is here.

In each instance, I concluded the planning conversation with a request for the LLM to produce precise instructions that I could then paste into Claude Code (or Codex) so that it could build the simulator. Opus 4.6 titled its implementation “The Dojo”, while GPT 5.4 went with “Project Asterion” — fitting titles given the approach of each respective simulator. The links above will take readers to the GitHub repos with the code and supporting documentation for each project.

The simulators

Readers can enter the Dojo by clicking this button:

You will need a free Anthropic API key (see the footnote for step-by-step instructions on how to get one).¹

Project Asterion consists of a bigger and more involved simulated world. To enter it readers need to clone the repo and run it locally (no API key needed).

Step-by-step instructions on how to do that are here.

I promise you, both are worth doing. (Plus, it is a great way for those new to vibe coding to learn the ropes.)

In the Dojo, the simulation includes a realistic bank, with a realistic VaR issue (HistSim, 12 month window), a boss who is knowledgeable but curt, and a realistic head of market risk (quantitatively strong, cards close to the chest). The interface is a little gamified, but that makes it surprisingly fun.

The first task in the Dojo eases you in. You get a high level description of the bank’s VaR model and are asked by your boss to select key weaknesses from a list of options — all of which sound plausible if you don’t know much about VaR. Once you make the selections, your boss gives you feedback.

Then you meet with the head of market risk. She asks if you have any questions. There’s a menu of suggested questions or you can type in your own question. In the screenshot below, I do the latter. Her response seems fine, but if you know VaR well you can detect some evasiveness. (Your boss’s note in the upper right warns: “Watch for what she redirects from.”)

The simulation somehow feels realistic and nuanced, especially considering that it was built in half a day.

The Project Asterion simulation is more comprehensive and less story-driven. Seasoned market risk examiners will likely find it more to their liking because of its coverage of practical exam skills and the scope of content. Below is a screenshot from one of the pre-reads before the first risk manager interview.

I asked both models to reflect on the differences between the two simulators. They agreed that the Dojo’s narrative structure is more engaging, while Asterion’s comprehensiveness substantively covers more of what junior examiners need to master. Opus 4.6 concluded:

The ideal build would merge both: GPT's substantive architecture and exam-process fidelity with Opus's NPC behavioral depth and the Phase 5 political pressure scenario. That's probably your next iteration.

Or yours, dear reader.

Hands on keys

To get proficient at AI, nothing beats “hands on keys”. I highly recommend tackling ambitious projects like this because they cover a lot of ground. Building a training simulator, in particular, accomplishes three things simultaneously:

1. It pushes you to use a coding agent and shows the value of extensive planning,

2. It gives you something substantive to review to get a sense of the strengths, weaknesses, and blindspots of today’s AI models, and

3. It builds a working prototype that has the potential to become practically useful for your team or agency. (It’s also fun.)

A quick note on technical requirements. It helps to have the $20/mo tier of the models, as that gives you access to the latest versions. Opus 4.6 and GPT 5.4, for instance, do noticeably better on reasoning, coding, tool use, factuality, etc., than what’s offered in the free tier. They were both able to look up exam procedures and VaR issues on their own, with no guidance from me. The free tier models aren’t as capable in those regards.

I also am a strong believer in feeding the same prompts and projects into multiple models, as that provides healthy perspective and prevents over-reliance on any given model.

In terms of coding agents, for me Claude Code in the terminal has been by go-to because it covers the last mile of coding quite well, e.g., connecting to a website that others can access via Substack, creating the repo on GitHub so the code can be shared, etc. After code gets written, some coding agents and interfaces assume you know how to make it go live or share it. I don’t. I’ve found that the terminal version of Claude Code will simply do it, rather than tell me how to do it, which is nice for a non-coder. (My guess is that other coding agents will catch up on this quickly.)

Looking ahead

In my post “The New AI Jobs in Supervision” I noted that as AI displaces certain tasks, it is also going to create opportunities for supervisors to be AI integrators, builders, and trust engineers. Developing training simulators like the Dojo combines all three. As importantly, it focuses us on the very real challenge of ensuring that junior talent has opportunities to learn and grow.

My hope is that this v0.1 experiment inspires others to tinker and build more sophisticated and professional grade training simulators. The Dojo and Asterion were spun up in a day in a half by myself. It’s exciting to consider what a team over a couple of weeks can build.

Finally, if you are a regulator who has been playing around with AI creating tools/agents and are looking for a community of like-minded builders to share ideas and code with, reach out to me on LinkedIn with an example or two of what you’ve created. You are not alone.

1. To get a free Anthropic API key:

1. Go to console.anthropic.com and create a free account

2. Once logged in, click “API Keys” in the left sidebar

3. Click “Create Key”, give it a name (e.g. “The Dojo”), and copy it — it starts with sk-ant-

4. Anthropic gives new accounts $5 of free credit, which is enough for dozens of exam sessions

5. Paste the key into the simulator when prompted — it stays in your browser and is never stored anywhere else