How Artificial Intelligence Helps in Writing Design Documents

Lead Game Designer Crazy Panda Sergey Zaigraev shared with App2Top how he utilizes neural networks for working on documentation.

This article was published with the support of the educational service WN Academy, which actively recruits leading experts for its courses and corporate training. 

Sergey Zaigraev

Hello! My name is Sergey, I'm a game designer.

I love reinventing the wheel. This time, I want to share my latest invention: a story about how I use generative language models for work.

You've probably already used neural networks and even achieved some success. I want to share my method and invite you to discuss it. I'd be very interested to know how other game designers use this tool.

First Relationship with Neural Networks

I started using neural networks as soon as they appeared — I immediately adapted them for small automation tasks and coding. Usually, these are short scripts, like my pet project for working with game configurations.

Of course, I also started drawing funny pictures and writing amusing texts. But usually, these did not extend beyond cool and entertaining stuff in a vacuum. They were hard to call practical and useful.

In other words, I initially couldn't use neural networks as a tool for work. All my requests related to game design were met with very generic recommendations, reminiscent of Captain Obvious.

Over time, I found several productive applications for AI, which we will discuss.

The first is assisting in writing Game Design Documents (GDD) by offloading the structuring and formatting of the text to the neural network. This allows me to focus on the mechanics and player interaction without being distracted by editing during the process.

The second is receiving feedback on described ideas and the completed document, which sometimes leads to an improved result.

From Procrastination to Task Setting

Writing a design document is a regular source of procrastination for me. Especially in the first hours, when I'm staring at a blank page trying to organize a whole heap of ideas into a coherent structure. Usually, all this doesn't fit into memory and I start to stall.

A few iterations later, the problem resolves itself, the document starts coming together and takes shape. However, it's a tedious process that I previously couldn't optimize. Not to mention that during numerous approvals and rewritings, in a document that's rapidly growing in size, it's easy to get lost.

Eventually, I realized I wanted a tool that would significantly simplify the process of writing a design document. I set the context, the general description at a high level, comment on the nuances of implementation, and let the neural network do everything else. By fixing details at the end, I can produce a GDD that isn't embarrassing to show the team, it's clear and concise. Working collaboratively with an LLM (Large Language Model) allows for writing more detailed documents, avoiding absurd mistakes due to inattention, paying more attention to mechanics, focusing on player interactions.

I formulated it like this for myself: I need a tool that:

• solves the problem of a blank page;
• resolves the issue of memory overload, when ideas need to be dumped somewhere reliable to unlock further thoughts;
• helps create the basic structure of the document;
• assists in the initial evaluation of mechanics when you've already become blind to them;
• helps with working out tricky cases that a negligent game designer might forget;
• allows focusing on the mechanics' specific features and their interaction with the player;
• enables concentrating on integrating the mechanics into a particular game.

For me, this tool has become an AI chat.

Currently, I primarily use the Chinese chat DeepSeek, which is accessible, free, quite smart, and not as bland in its responses as other neural networks.

4 Stages of Working with a Neural Network When Creating a Design Document

I have divided further work with the chat into several stages.

First Stage. Set the Context

When I want something from other people, I need to clearly explain what exactly I want. The better the explanation, the better the result.

The same goes for AI: the more and clearer the context, the better defined the task, the better the outcome.

To do this, I open a new chat and explain to DeepSeek that it is my personal assistant, and its task is to organize my notes and maintain order.

I use a prompt like this:

"You are a digital secretary who transforms voice notes from a 'stream of consciousness' format into structured drafts for documents or articles. Your task is to preserve the original author's style, highlight the logic, and organize thoughts without adding interpretations, advice, or creativity."

Following this, I provide the next set of information.

The prompt doesn't necessarily have to be this complex; you can eliminate half of it, and it will still work. AI helped me write the prompt, by the way.

Second Stage. Unloading the Stream of Consciousness

Now I can pour the stream of consciousness onto the prepared ground. For this, I open the DeepSeek app on my smartphone and start recording it with voice messages through the voice input keyboard. It's literally a stream of consciousness: ideas, notes, interesting implementations from other games, just cool ideas that might be fitting.

I don't even try to follow a specific structure; I speak in a random order, pile ideas up. I unload the contents of my head and notes onto the AI. Everything collected needs to be put into the chat.

It might look something like this:

"we need to describe the mechanics of enhancing the viewing of rewards for the season pass. the main change we're making is now everything appears beautifully on panels, borders for avatars are removed, and when tapping on panel rewards like gifts, chests, achievements slots for them open with a detailed description, the box button contains information about guaranteed rewards and what he can get with a chance. for the trophy, the trophy itself opens along with a description of where it can be obtained. for the gift, an image of the gift appears with the caption, and for items, a large preview with the player's avatar as they would look in the store. if it's a costume, the avatar is seen full-length, if a headpiece, then the head is in close-up, exactly how it is in the store, meaning the camera positions like in the wardrobe when choosing items depending on the category of the item".

The Google voice input doesn't always work well and completely ignores punctuation, so try to highlight accents with words. Even such a stream of consciousness can be parsed and organized by AI.

Speaking, in fact, is very beneficial on its own. Personally, as I speak, new ideas always begin to appear. This seems to be the key to this method.

Third Stage. Describe the Game and Its Connection with the New Mechanics

With the ideas done, now it's necessary to describe the game. What kind of game it is, what is its audience, what mechanics exist, how these mechanics interact with each other. Without fanaticism, but all the main points need to be listed in the voice chat.

After that, you need to describe the expectations from the new mechanics: how they affect the player, what they're aimed at, how they're related to the existing mechanics.

Both the first and the second should be described as detailed as possible. Don't be lazy. The more information, the more adequate and relevant the response will be. These are the frameworks and conditions modulating the neural network's response.

Undoubtedly, AI will still add something of its own, make suggestions and assumptions. But first, with well-defined boundaries, this will be less than usual. Second, everything will relate to the specific project, which can be a plus.

It's also worth mentioning that the more structured the stream of consciousness is, the better, but to start, just capturing ideas is enough. As you speak, they can organically form a structure.

Fourth Stage. Organize

At this point, we have:

• thrown in ideas;
• explained the project;
• specified the connection of the project with the new mechanics.

That means it's time to put things in order. For this, we ask the AI to create a structure from all the downloaded information. We request it roughly like this:

"You are an experienced game designer and technical writer specializing in creating clear, structured, and useful design documents. Your task is to help me transform my scattered ideas, sketches, and thoughts into a professional document understandable to every member of the development team.

For this:

• • create a logical structure for the document with sections and subsections;
  • systematize my ideas and distribute them across relevant sections;
  • highlight important points that require further elaboration;
  • point out any possible contradictions or illogicalities in my ideas;
  • suggest additional ideas or improvements where appropriate;
  • formulate clear questions where there isn't enough information for decision making."

Usually, it doesn't work on the first try. The response will have too much superfluity or be too generic. For this reason, I often return to the last message with instructions and tweak it: I describe desired behavior and specify what not to do. You can guide the AI on the structure to stick to.

Through several iterations, a coherent result is achieved. Returning and adjusting the request is a fairly universal rule. Every time AI starts deviating, the most effective approach is to edit the prompt, taking into account the neural network’s response.

This is literally like a roguelike! Make another run, based on AI’s responses in previous runs.

What to Do if AI Starts Confusing

At some point, AI will nevertheless get confused with corrections, start forgetting details, generate nonsense, and completely divert. In such a case, one should:

1. take the latest result as is and start editing manually (at this stage the document already gains a coherent form);
2. send the manually corrected document into a new chat with a note that it is not final;
3. continue adding ideas in the new chat;
4. ask to integrate them into the document structure;
5. repeat step 4, if it doesn't help, back to step 1.

Somewhere here, I usually reach the limit of the tool’s effectiveness. The document is already quite structured, most of my new ideas automatically fit into this structure, and working with the document in chat becomes inconvenient. Sometimes it takes longer to explain than to do it manually. It’s crucial to catch this moment and copy the document to a text editor and continue working with it independently.

Important: even if the document seems good, it still needs thorough proofreading and editing. You shouldn't hand it over to the development team straight from the AI.

Generally, by this point, you will have a solid concept document with a clear description of the game idea or its mechanics. It is open for discussion with colleagues and for further work.

AI for Feedback

In the process, I regularly upload working versions of the document to AI, asking for criticism and suggestions on different segments. Specific questions or doubts about particular text fragments are often quite interesting and useful. I recommend experimenting with existing documents and mechanics.

Out of curiosity, try submitting your completed documents to AI with questions like:

• how else can this be done?
• how is this done in other games (provide examples)?
• give advice on such-and-such a mechanic, considering how it interacts with such-and-such a mechanic;
• evaluate from the perspective of a player who (enter player style description).

You can also try the following amusing experiment:

• ask AI to make a summary from the completed document;
• submit this summary back to AI with instructions to create a design document.

Sometimes, interesting results emerge.

Conclusion

For me, AI made the process of writing documentation easier and definitely more enjoyable. Starting to speak as it is — is very simple, plus it quickly immerses you into work. Within minutes, I'm fully engaged in the process, even if I was staring at a blank page a moment ago.

Yes, the tool isn't the most convenient, but it is available to everyone and solves the tasks set at the beginning of the article.

Also, with the advent of neural networks, my code has noticeably improved! Haha.