How Artificial Intelligence Assists in Writing Design Documents

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

Sergey Zaigraev

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

I love reinventing the wheel. This time I'd like to share my next invention: a story about how I use generative language models in my work.

You've probably used neural networks and even achieved some success. I want to share my approach and invite further discussion. I would be very interested to learn how other game designers use this tool.

First encounters with neural networks

I started using neural networks as soon as they became available—almost immediately adapting them for minor automations 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, in my case, this didn't go beyond cool and entertaining things in a vacuum. It was difficult to call them practical and useful.

In other words, initially, I couldn't use neural networks as a tool for work. To all my game design inquiries, I always received very general recommendations in the style of Captain Obvious.

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

The first one is assistance in writing GDD by delegating the structuring and formatting of text to the neural network. This helps focus on mechanics and interaction with the player, without getting distracted by editing.

The second one is receiving feedback on described ideas and the finished document, sometimes leading to an improved outcome.

From procrastination to task setting

Writing a design document is a regular source of procrastination for me. Especially in the first hours, when I sit in front of a blank page trying to fit a heap of ideas into a coherent structure. Usually, this doesn't fit into my mental RAM, and I start to stall.

A few iterations solve the problem, the document starts taking shape and becomes tangible. However, it's a tedious process that I couldn't optimize. This doesn't account for the maze of approvals and rewrites in a document rapidly growing in size, making it easy to get lost.

Ultimately, I realized I wanted a tool to simplify the design document writing process. I set the context, provide a high-level overview, comment on implementation nuances, and let the neural network handle the rest. By tweaking details at the end, I'll get a GDD I'm not ashamed to share with the team—a clear and understandable one. Collaborative work with LLM allows more elaborated documents, avoiding silly mistakes due to inattention, and focusing more on mechanics and how they interact with the player.

For myself, I formulated it as needing a tool that:

• solves the blank page problem;
• addresses memory overload, unloading ideas into a safe place for unlocking further thoughts;
• helps create a basic document structure;
• assists in initial evaluation when your eyes are tired;
• helps work through tricky cases forgotten by an inattentive game designer;
• focuses on the mechanics' features and their interaction with the player;
• focuses on integrating mechanics into a specific game.

This tool became an AI chat for me.

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

4 stages of working with neural networks in creating a design document

I divided the chat work into several stages.

First stage. Set the context

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

The same goes for AI: the more detailed and clearer the context, and the better the task is set, the better the result will be.

For this, I open a new chat and explain to DeepSeek that it is my personal assistant, and its task is to keep my notes organized.

I use a prompt like this:

"You are a digital secretary that transforms stream-of-consciousness voice notes into structured document drafts or articles. Your task is to maintain the author's original style, highlight logic, and organize thoughts without adding interpretations, advice, or creativity."

After this, I provide the following information.

The prompt doesn't have to be this complex; you can remove half of it and it will still work. Incidentally, I got help from the AI with writing the prompt.

Second stage. Unload the stream of consciousness

Now, I can pour my stream of consciousness onto the prepared platform. To do this, I open the DeepSeek app on my smartphone and start recording voice messages through the voice input of the keyboard. This is literally a stream of consciousness: ideas, notes, interesting implementations from other games, just cool ideas that might fit.

I don't even try to follow any structure; I speak in random order, tossing ideas into a pile. I unload the contents of my head and notes into the AI. Everything accumulated needs to be transferred into the chat.

It might look something like this:

"we need to describe the mechanic of improving reward viewing for the season pass. the main change we're making is everything is now displayed beautifully in panels removed borders for avatars and on tapping a reward panel like gifts prize boxes achievements slots for them a window opens with a detailed description box button contains info on guaranteed rewards and what can be obtained by chance. for the trophy its own description opens with information on where it can be obtained. for the gift shows a picture of the gift and title of gift and for items a large preview open with player's avatar as it would appear in the store so if it's a costume you see the avatar full-length if it's a headdress you see head close-up exactly as in the store so the camera is positioned just like in the wardrobe when selecting items depending on the item category".

Google's voice input doesn't always work perfectly and ignores punctuation, so try to highlight accents with words. Even such a stream of consciousness, the AI can sort and organize.

As a side note, articulating is quite helpful on its own. Personally, while speaking, new ideas often emerge. This seems to be the key part of the method.

Third stage. Describe the game and the connection with the new mechanic

With ideas covered, now it's time to describe the game. What game is it, who is the audience, what are the mechanics, how do these mechanics interact? Not fanatically, but listing all the essentials in a voice note to the chat.

Next, describe expectations from the new mechanic, how it impacts the player, what it's aimed at, how it relates to the existing mechanics.

Describing both should be 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 shaping the neural network's response.

Of course, AI will still add some of its own, make suggestions and assumptions. However, firstly, with well-defined boundaries, there will be less of this than usual. Secondly, everything will relate to a specific project, which can be a plus.

It's also worth noting that the more structured the stream of consciousness, the better, but for beginners, just capturing ideas is enough. During the articulation process, they might naturally organize into a structure.

Fourth stage. Organize everything

By this point, we have:

• thrown out ideas;
• discussed the project;
• outlined the connection between the project and new mechanics.

This means it’s time to organize. To do this, ask the AI to structure all the information. Something like:

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

For this:

• create a logical structure for the document with sections and subsections;
• systematize my ideas and distribute them to relevant sections;
• highlight important points that require further development;
• point out potential contradictions or illogicalities in my ideas;
• suggest additional ideas or improvements where appropriate;
• formulate clear questions where there is a lack of information for decision-making."

Usually, it doesn’t succeed on the first try. The response might be full of unnecessary details or overly general. Often, I go back to the latest message with instructions and adjust it: specify desired behavior, clarify what not to do. You can also suggest to AI what structure to adhere to.

Through several iterations, you can achieve a sensible result. Going back and modifying the request is a fairly universal rule. Whenever AI starts making independent decisions, the most effective way is to edit the prompt, taking into account the neural network's response.

It’s literally like a roguelike! Make another run, based on the AI's responses in previous runs.

What to do if AI starts getting confused

At some point, AI will inevitably get tangled in edits, forget details, and generate nonsense or completely diverge from the subject. In such a case:

1. take the latest result as it is and start editing manually (by this stage the document is already gaining a coherent form);
2. send the manually edited document into a new chat, marking that it’s not final;
3. continue throwing new ideas into the new chat;
4. ask to incorporate them into the document's structure;
5. repeat step 4, if it doesn’t help, return to step 1.

Somewhere here, I usually hit the tool’s efficiency limit. The document is already sufficiently structured, most of my new ideas automatically fold into this structure, and working with the document in chat becomes inconvenient. Sometimes explaining takes longer than doing it yourself. You must sense this moment, copy the document into a text editor, and continue working with it independently.

Important: Even if it seems that the document turned out well, it still requires a careful read-through and adjustment. It can't be handed to the development team right after AI.

Typically, at this point, you have a reasonably good concept document that clearly describes the game idea or mechanics. It can be discussed with colleagues and used for further work.

AI for feedback

During the work process, I regularly upload working versions of the document to AI, ask for critiques, and suggestions on different blocks. Specific questions or doubts about particular text fragments are often quite intriguing and useful. I recommend experimenting with existing documents and mechanics.

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

• How else could this be done?
• How is this done in other games (provide examples)?
• Give advice on such-and-such mechanics, considering how it interacts with so-and-so mechanics;
• Evaluate from the perspective of a player who (description of player style here).

Another fun experiment to conduct is:

• ask AI to summarize a completed document;
• feed this summary back to AI with instructions to make a design document.

Sometimes this yields interesting results.

Conclusion

For me, AI made the documentation writing process simpler and unquestionably more enjoyable. It’s very easy to start talking as things are, plus it quickly draws you into the work. Literally in a few minutes, I am fully immersed in the process, even if I’ve just been staring at a blank page.

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

Additionally, with the advent of neural networks, my code has become notably better! Hehe.