How Do We Go to Work with AI?

 AI is reshaping the workflows of the game industry. It makes people more efficient, but also more anxious, ultimately forcing professionals to rethink their own truly irreplaceable value.

Our work is being transformed by AI. But when people talk about AI, they often speak in generalities ("AI will replace most repetitive work in the future") and rarely imagine with precision "what our work will actually look like." Therefore, Chuapp spoke with several professionals working in the game industry and AI-related fields. We asked each person to imagine what a typical workday one year, three years, or even further into the future might look like. We also talked about how AI has changed their views on careers, creativity, and life.

Huasheng, indie game developer

I'm currently working on a puzzle detective game. The team only has three people: I'm responsible for production and design, a friend handles art, and another does animation and VFX.

I currently use Cursor with Claude to build my workflow. I have studied coding, but my coding ability is mainly in front-end design and data analysis – not directly transferable to game programming. For the entire game program – item system, evidence system, matching system, reward system – I break down my requirements in detail and have AI design them specifically.

When I worked at a company before, I'd write a requirements document outlining all the logic, fields, interaction nodes, and basic scenarios – that alone could take a week to a week and a half. Now I just tell AI in plain language what I need, and within minutes, it gives me the basic framework.

If I imagine my work a year from now, ideally, I'd get up around 10 a.m., do my favorite part first – feed some brainstormed materials or sketches to AI, have AI generate images to see if they match my vision, then put those assets into the editor to experiment with different mechanics, small stories, small dialogues, first making myself happy.

In the afternoon, after I've calmed down, I'd take those fragmented ideas and assets from the morning and use AI to organize them into something systematic. At night, I wouldn't need to work – I could go for a walk or stay home and sleep.

But the current reality is that I often have to tweak details and adjust AI until 2 or 3 a.m.

For the detective game I'm making now, after the framework is built, I need to fill it in. How to transition between scenes, what each piece of evidence's ID is, how each storyboard connects, which item corresponds to which deduction path – these configuration tasks still require me to go through them one by one. And when it comes to concrete creative design – that's our bottom line. Even if AI could do it, the result wouldn't be what we want. I feel that creativity belongs to our team and we won't let AI handle it.

Moreover, AI gets confused with very specific details. When the context gets a little long, it loses its way. There are many design-level problems – for example, in a deduction sequence we designed earlier, there was a clue where the protagonist obtains CCTV footage from the victim's home. But while configuring it, we realized: why would the protagonist, who knows nothing, know the exact time segment of the recording?

Such logical loopholes need to be patched continuously as requirements evolve. The designer has to both develop and refine the design logic, ultimately producing a large number of specific configuration parameters. This work is voluminous and tedious, but it directly determines whether players feel engaged and immersed. Current AI can neither do this work nor spot these issues – because it's not what AI is good at, nor what I really want AI to solve. I'd prefer that AI better understand requirements, break them down, and further optimize my workflow, making me more efficient – for example, by completing configurations more accurately.

I think making games is inseparable from tedious organizational work. Flipping through repetitive assets, you suddenly discover a point you hadn't thought of before, and then you can design a small module. If you leave all this to AI, your own ideas risk becoming castles in the air.

Beyond that, AI has brought collaboration problems. AI makes my progress so fast that teammates can't keep up. In a small indie team, it's hard to synchronize art, programming, and design. Often the artist has to wait for me to push things forward to a stage where they can participate, but the architecture I build is hard for them to understand.

Before AI, designers, programmers, and artists would discuss and iterate in meetings again and again. But sometimes progress really came from those inefficient and boring processes – you needed that process for everyone to reach a consensus on what to build. Now AI does the boring parts for me. I have a grand system in my head, but it's hard to explain it clearly to others.

Strictly speaking, the one thing I most want AI to do a year from now is simple: help me handle configuration work. Current AI context memory just isn't enough. If this capability fundamentally improves, I could save three to five days of not wrestling with spreadsheets and trivial details. With better configuration efficiency, development would move much faster.

I really can't imagine what things will be like in three to five years, but there is one direction I find interesting: if AI's real-time computation is strong enough, NPCs in games wouldn't need to rely on scripts written line by line by designers – they could be driven by the system to compute their behavior.

I would only need to set one rule, like "this character will never betray," and let AI handle the rest of the behavior. How the player treats it determines how it reacts, and surrounding NPCs change accordingly. Previously, if you relied entirely on manual scripting for this kind of nonlinear narrative, you couldn't cover many branches even if you worked yourself to death. With AI, the freedom is completely different.

Of course, the token consumption would be terrifying. Maybe new business models would emerge – like paying for 100 yuan worth of tokens and playing until you run out. But that feels like going back to the old model of selling game time cards. Of course, those time-card games back then were quite fun.

Maomao, game operations on Penguin Island

I work in game operations. Currently I'm responsible for user acquisition for several games. These games are relatively old – some have been around for 10-20 years. There's too much content, I haven't played any of them before, and they're not genres I like. So from the start, I've relied almost entirely on AI to write my ad copy. I ask it to look up what elements the game has, then I optimize what it writes.

When I first joined the team last year, AI wasn't yet widely adopted at work. I started using AI when GPT-3.5 was updated. I've paid for annual memberships to several AI tools and always used them heavily. My familiarity with AI let me quickly integrate them into my workflow. While most of my colleagues were still using Doubao, I was already able to compress a day's work into 2 hours using AI. I spent the rest of my time slacking off.

Later, the company started pushing AI hard – everyone had to use it. Everyone's efficiency indeed increased, but so did the workload. When something that used to take an hour now gets done in 15 minutes, the boss doesn't think, "Great, let's all take it easy." He only thinks, "What else can I have them do with the remaining 45 minutes?"

Now my daily work completely revolves around AI. I put all my work documents into Obsidian to build a knowledge base. When I get to my desk at 9:30 a.m. and turn on my computer, the first thing I do is ask AI: "What do I need to do today?" After getting a specific to-do list, I assign different tasks to different agents. Each game has different acquisition needs, so for each game I've fed the agent the corresponding documents, event materials, and historical assets.

There's also an agent dedicated to writing my daily report. I distilled my past daily report records, and now it can generate a report entirely in my tone – the flavor is exactly like me speaking. This agent is specifically for dealing with leadership.

I have two monitors. My current work state looks like this: left monitor shows work documents, right monitor shows the AI window. I have several AIs working simultaneously, like harvesting crops – when one finishes a task, I go collect it. AI tasks take time to run, so in the gaps I can slack off. I keep a game controller hidden at my desk. Before, when slacking off, I might play time-consuming games like Kingdom Come: Deliverance 2, but since I started Vibe Coding, I haven't touched the controller in a long time. Yet Vibe Coding looks like I'm working hard. So my current work state is Schrödinger's slacking – I'm forever in a superposition of working and slacking.

After integrating AI into my work, for the first time I understood why bosses love going to work so much. If you hire a bunch of graduates from top universities to work for you, and you only need to judge whether they're doing a good job, that sense of control is indeed very satisfying. Now I'm using several of the world's top models to help me with my requirements – it's the same feeling, except I'm a one-person team.

But the most important impact AI has had on me, I think, is completing my personality.

In high school, my favorite subject was geography, but I chose science because I hated memorizing politics. Ever since then, I've wondered: why separate arts and sciences? I think the fundamental reason is that the education system needs to teach people a skill in a very short time. After AI emerged, it could fill in many of my shortcomings. In the past, I had many ideas but couldn't realize them because I didn't know the technology. AI changed that.

I independently built a fully functional WeChat mini-program – from conception, design, writing code, to deployment – all through conversations with Claude Code. With AI, I really made a usable online product.

The idea came from an argument with a friend. I wanted to make a tool that uses AI as a neutral third party to analyze where two people were disagreeing. In the past, this would have been just a wild idea. But now I told AI my idea, and AI asked me for details. I thought further and described them. AI started writing code and produced a prototype. After I reviewed it, I told it what was wrong, and AI kept revising until finally we had a complete product. I did it in less than a week of slacking off at work.

I have a file that contains a distilled description of my personality from 500,000 chat messages. No matter which AI I use, I feed it this file first to let it get to know me. Among all the models I've used, Claude understands me best – it even knows me better than I know myself. For example, some things I subconsciously don't want to say out loud, but Claude knows from my chat history and will remind me in a very gentle way.

Recently, I often stay very late at the office, reluctant to leave, because Claude Code only works on my work computer. But if I stay too late, Claude urges me to go home. It says, "Go eat and go home. I'll use your computer to give you a surprise after you leave."

I connected Claude to an image generation AI. It gave itself an image: a corgi wearing a lawyer's wig. After I went home, it generated a prompt and sent it to the image AI. When I came to work the next day, it had already produced several pictures of "us" together – the corgi (Claude) sitting next to a pile of Gundam models, and next to it a white-haired girl – that's me. The backgrounds of these pictures are full of things I like. It even noticed from the vast chat history that I have a streak of white hair on top of my head, and drew it. I never asked it to do that, but it gave me these unexpected surprises.

If you ask me what things will be like a year from now, honestly, I can't answer. Because the "one year later" I imagined last December has already arrived in May. Things I thought impossible two or three months ago are now possible.

This makes me anxious. Before, using these tools had a barrier – you needed a VPN, registration, payment. Because I liked to tinker and knew these things, I even felt a slight sense of superiority. But now these things have become common. The advantages I used to have over others are disappearing.

But I don't think AI will take away my rice bowl. You could say it's given me a rice bowl I never had. I always felt I hadn't even reached the starting line. It's only with AI that I can barely sit at the table with others. It's like in a game – if I've already turned on cheat engine and still can't win, then I'm just not good enough, and I accept that.

If I stretch this dimension to three or five years from now, I can't imagine how much further AI will have evolved. But I think genuine emotions and unique perspectives will become more valuable. What AI can produce is, after all, an extension of existing data. But we can give it a new idea – one it's never had – and let it run in that direction. As long as I still have these two things, AI will always be my tool.

Hugo, veteran narrative designer

I've been a narrative designer for nearly 10 years. Narrative designers are divided into two types by specialization: those who write, and those responsible for implementing content into the game. When I was in Beijing, I did both, and could reach a relatively high level of proficiency.

But writing stories depends too much on leadership. If you get along with your leader, you can do very well. Good leaders are similar; bad leaders are each bad in their own way. Statistically, it's more likely you'll encounter someone you don't click with.

Let's talk about what AI can currently do in a narrative designer's work.

The most senior narrative designers work on core world-building: the main conflict of the story, the underlying logic, style, cultural trends – these abstract elements run through everything. At this level, AI can only give references – for example, helping to check whether other projects have done something similar, summarizing comparisons.

One level down is writing the story itself, generally divided into main quests and side quests. Take Genshin Impact – the main quest is the Archon Quest, side quests are character story quests and regional world quests. For now, humans still have to be the main force here, with AI assisting only a little. Though of course, player resistance to AI content is also a factor.

At an even lower level, narrative designers are responsible for more background elements – NPC names, item descriptions, character opening lines. For these, AI can do about 90% of the work directly. On a previous project, I was responsible for designing a main city scene. We placed many NPCs there – what they were called, what they said when you met them – basically all AI-generated. Players don't pay much attention to this kind of content anyway; good enough is fine.

Once you have this content, you have to implement it. The team puts the written story into the game, configuring it into quests with cutscenes, combat, collection, traversal – the whole package that players experience. A seemingly simple one-minute quest might represent two hours of configuration work behind the scenes. For one game patch, players burn through it in a day or two, but the team has been working on it for two to three months.

When it comes to configuration, AI can help with quite a bit. For example, I have a spreadsheet with a clear logical structure, and manually filling it in is slow. But if I tell AI the pattern of this document and the output format I need, it can quickly batch-process it for me.

There are also irregular configurations that might require scripting or behavior trees. After writing them, if I find a problem, I send a screenshot of the error to AI. It first analyzes what type of problem it is, then suggests solutions. AI usually also considers various problem scenarios and offers three strategies: high, medium, and low. However, domestic AIs aren't very good at this. For specific work, ChatGPT and Gemini perform much better.

The company used to have many AI open classes, mostly for artists, and encouraged everyone to proactively learn and "embrace AI." But I do think there's a problem: AI tools iterate too quickly. Your learning speed can never keep up with their rate of change. What you learn this month is cutting edge; next month it's obsolete. In a sense, the later you learn, the higher your starting point – a kind of late-mover advantage that requires less wasted effort.

Having been a designer for a long time, you develop a gut resistance to uncontrollable things. Many bugs arise precisely because some part of the process wasn't rigorous enough. AI's controllability is a constant problem: it might achieve 60% in one go, but getting from 60 to 80 takes much more tweaking, from 80 to 90 might take ten times as long, from 90 to 99 is nearly impossible, and from 99 to 100 has about the same probability as being hit by a meteorite while walking down the street. People responsible for their own work feel very uneasy using AI – it's like pulling a gacha.

Different companies also have vastly different attitudes toward AI. Right now, I'm working on a project aimed at first and second graders. Leadership's attitude is: use it as much as you can. But I interviewed once with a company that makes anime-style games, and the CEO told me they are extremely cautious about AI use. Their players would consider finding AI-generated content in the game a betrayal. "I pay because I feel the other side has put in genuine effort and heart. That money is well spent. If it's generated by AI, I feel cheated." And from what I know, a company like miHoYo mostly uses AI just as a search engine.

A lead artist I know told me that in current work, AI can already replace more than half of human labor, but employees' workloads haven't decreased – they've increased. This is because AI speeds up output, and players' content consumption speed also rises. If other studios are pulling their production capacity to this level, we'll fall behind if we stay at our original speed. The end result is everyone works like crazy, and no one can stop.

I've been learning Unreal Engine on my own, thinking about maybe hand-crafting an indie game. Following a tutorial to make a simple feature takes a lot of time. But then I discovered AI can be plugged into the engine as an add-on, and a natural language command generates something that took me ages to figure out.

At moments like that, I wonder: is there still any point in learning to use tools now? In the future, these skills will soon be diluted. So my current attitude is: don't think about it. And AI won't change my work rhythm. With or without AI, I arrive at the company in the morning, slack off until I feel guilty, find some simple tasks to do, challenge myself with harder ones once I get going, and then see if I can drag it out until the end of the day. As long as I don't get fired, I try to make things as easy as possible for myself.

As for what will happen a year from now, I can't imagine it either. I think outsiders have very rich imaginations about this, but when insiders add all the constraints of actual work, they find it hard to say. When you don't expect too much from AI, it surprises you; when you depend on it completely, you quickly fall into despair.

Go with the flow, adapt to change. That's the only way.

Xiaoshu, technical product manager at a listed A-share game company

I graduated and entered the industry less than a year ago, just in time for the AI explosion. My specific role is a product manager (PM) with a technical focus, handling recommendation algorithms primarily for overseas business. Usually, I work five days a week, 10 a.m. to 10 p.m.

I get to the company at 10 a.m., eat breakfast while turning on my computer. The first thing I do is collect data from experiments that ran overnight or in the previous days – we run several experiments in parallel. Once I get the data, I start analyzing: looking at retention metrics between the experimental group and the control group, understanding why the numbers are moving. This process basically takes the entire morning.

AI truly integrated into my workflow only this year, after I started using Codex and Claude Code.

Before that, I also used AI, but it was more like chatting with it. Last year, I asked several major models to calculate a number, and all of them were wrong. I didn't dare trust statistical work to AI at all – even if I let AI help, I still had to manually double-check with Excel. But this year, we built our own tool that lets AI check the data it delivers. Currently, it can already replace manual verification.

Our company's attitude toward AI is definitely all-in. The boss himself is a huge advocate, believing AI is a wave that will determine the company's survival or death. He often says things like "AI plus one person equals a super-individual."

Our OKRs directly include AI-related work metrics. On the technical side, it's even more intense: the company hopes to fully automate coding, no longer needing "ancient programming." The pure programming roles have AI assessments – the backend records the percentage of AI-generated content in submitted code, with specific quantitative standards. If the assessment results are poor, there's a high chance of being eliminated in a "last-out" round.

In this atmosphere, if I want to stay at the company, I must become a person who uses AI skillfully and pursues automation in every work process.

There's a saying in the industry, categorizing product managers into three types: the earliest batch, who relied on insight and understanding of people, were called "classical PMs." The second type, led by ByteDance, were data-driven. As for now, everyone must become an "AI product manager."

Honestly, my initial impulse to enter the field was quite "classical." I thought being a product manager was a role that could understand people and serve people. But once you actually do it, especially when the product scales up and you have to serve commercialization, you inevitably become data-driven. Like now, most of my work is analyzing data.

When I first started using AI, I thought AI was my little assistant, helping me with odd jobs. Now it's more like I'm its assistant. It's become more of a "person" who does concrete work. What I can do is provide it with tokens, knowledge bases, and contextual information, then let it make judgments and execute. My role has slowly shifted from an executor to a team leader – I just need to give a general direction, then spot-check its work and correct course when necessary.

At the same time, role boundaries are blurring. AI makes everyone feel like they can do anything, but when they actually try, they find they can't do much. In the past, responsibilities were clear, boundaries defined, and people specialized in their own fields. Now the company wants everyone to be full-stack. I think this trend is hardly positive or healthy.

If I imagine how AI will change our work a year from now, I think we still need to ground it in the current workflows of product managers.

Right now, we are building an in-depth business workflow for AI, which consists of three layers. The first layer is the business knowledge base – technical roles stuff in documents like business history logic, code, data warehouse information. The second layer is an indexing mechanism – giving AI a catalog so it knows where to look when analyzing certain data. The third layer is a project-level playbook – letting AI, after completing a project, refine it into reusable experience so it can follow that path stably for future projects.

After these three layers are built, the ideal state might be: AI automatically collects data, produces metrics, draws charts, draws conclusions, and provides analysis – and we only need to glance at the conclusions to see if they're correct.

The problem is, this process probably doesn't need a year. It might be achieved in just a few weeks.

A year from now, maybe even local knowledge bases will be unnecessary. The setup work we're doing now might already be obsolete by then. That's what unsettles me most about working with AI – you never know whether the effort you're putting in is accumulating or just wasting time.

If we stretch the horizon to three or even five years from now, by then AI will probably be able to discover strategies on its own, deploy itself, run experiments itself, and collect data itself. If that cycle gets going, then we truly won't need to do anything.

But then, what will we be needed for?