Before the Build: Learning What AI Could—and Could Not—Do for RPGs

Part of today’s story reaches back to 2022 and 2023.

Before this year’s renewed push on P.A.T.H.: Vietnam, I spent a concentrated period experimenting with artificial intelligence and tabletop role-playing. I documented some of that work on ChatGPT RPG, a short-lived blog that remains online. I tested AI-guided solo and cooperative play, prompt structures, game-master behaviour, character creation, campaign augmentation and early combinations of writing, image and animation tools.

I also worked on a ChatGPT RPG guide and published colouring and activity books. My Amazon author page preserves part of that publishing period. Those projects were smaller and very different from P.A.T.H.: Vietnam, but they taught me how an idea travels from a prompt or rough page into an object somebody else can actually open.

They also taught me not to confuse generation with design.

The exciting part was immediate

Early language models could take a few details and return a character, location, encounter or branch in seconds. For solo role-playing, that felt like opening a door. A player who had no group available could ask for a scene. A game master could request alternatives when preparation stalled. A designer could test whether a premise produced interesting questions before spending a week polishing it.

Image tools added another layer. A character who had existed as two paragraphs could have a face. Animation and voice services could turn that portrait into a rough presentation. The result was not automatically good, original, accurate or ready to publish, but it made an idea easier to examine.

That was the real value: not magic completion, but fast visibility.

The limits arrived just as quickly

An AI game master could sound confident while forgetting an important fact from ten minutes earlier. It could offer several choices that were secretly the same choice. It could reward every plan, avoid difficult consequences or invent a rule when it did not know the answer. It could produce a great paragraph and then undermine it with the next one.

Long campaigns exposed the weakness. A role-playing world needs memory, state, rules, pacing and judgment. It needs to know whether an item was used, whether a person is alive, whether a promise was made and whether a consequence has already occurred. A language model improvising one response at a time does not automatically provide that structure.

The early experiments helped me separate two jobs that are still easy to mix together:

• generating possibilities;
• deciding what is true in the game.

The first can be assisted. The second requires authorship, evidence, rules and responsibility.

Prompting was not the same as directing

I learned that a longer prompt was not always a better prompt. Adding instructions could make the response sound more specific while hiding contradictions. The stronger approach was often iterative: define the task, ask for a small result, inspect it, identify what failed and revise the underlying structure.

That sounds obvious now. At the time, the speed of the tools encouraged the opposite behaviour. If one result was disappointing, it was tempting to generate ten more. Quantity created motion, but not necessarily progress.

The better habit was to stop and edit.

What did I actually want from this scene? Which facts were supplied? Which claims needed checking? What part was mine? What should never be delegated? If a result could not be explained, corrected and maintained, it was not yet useful game content.

Publishing made the lesson physical

The guide and colouring-book work added practical constraints. Pages needed dimensions. Images needed sufficient resolution. Repetition became obvious when several pages appeared together. Covers, titles, descriptions and categories had to make sense to a person who had not watched the experiment happen.

An AI-assisted idea still required selection, cleanup, layout, proofing and a decision about whether it deserved to exist as a product.

That experience matters to P.A.T.H.: Vietnam because the project already contains many kinds of material: prose, rules, research, images, records and artifacts. A pile of individually interesting outputs is not a coherent game. Every part has to belong to the same person, date, place and design intention.

The foundation I carried into P.A.T.H.

By January 2026, I no longer viewed AI as an automatic game master waiting to run a finished world for me. I saw it as a collection of assisted creative and engineering tools with very different strengths.

It could help me:

• compare structures and identify missing cases;
• turn a rough concept into several testable directions;
• examine a rules explanation from a new player’s perspective;
• draft code, tests or content for review;
• organize research questions and trace contradictions;
• create visual candidates that could be judged against an art direction;
• and challenge me to explain why one version belonged while another did not.

It could not decide the historical truth, own the creative intention, guarantee originality, remember a full campaign without engineered state, clear rights, or tell me when the project had become emotionally or physically unhealthy.

Those boundaries are becoming part of the design itself. I do not want an uncontrolled narrator inventing canon in front of the player. Content can be assisted during development, but any released experience will need authored rules, durable state, testing, provenance and human approval.

An old blog as a useful record

ChatGPT RPG now feels like a time capsule from a very early stage of public experimentation. Some ideas aged well. Others show how quickly the language around AI changed. I do not need to pretend every early conclusion was correct for the work to remain valuable.

The archive proves something simpler: I have been testing the relationship between AI, storytelling and role-playing for several years. Those tests give me both enthusiasm and caution as I decide what P.A.T.H.: Vietnam should become.

The enthusiasm helped me imagine a solo game with unusual depth.

The caution taught me that depth would have to be designed, recorded and reviewed—not merely requested.