YouMind
Sign in

I Built a WeChat Mini-Game Using ChatGPT

@songsong
SIMPLIFIED CHINESENov 18, 2026
516K
911
106
41
2.0K

TL;DR

A game developer details the process of creating a WeChat mini-game using only ChatGPT, sharing specific prompts, workflow adjustments, and techniques for generating levels via video analysis.

Over 300 WeChat mini-games generated more than 10 million RMB in revenue in a single quarter.

This data was released at the May 2026 WeChat Mini-Game Developer Conference. Also announced: over 80 mini-games have daily active users exceeding 1 million.

Everyone says the mini-game market is saturated, but seeing these numbers, I still wanted to know: Is there an opportunity for individuals to enter this business using AI?

I work at a mobile game company developing a well-known IP MMO. One of our channels is a mini-game version, so I know heavy games require massive manpower and resources.

But what about light games? Not SLG, MMO, or MOBA, but simple tap-and-swipe mechanics?

With that question in mind, I rolled up my sleeves and built a casual puzzle game called "Puzzles Must Be Solved" entirely using ChatGPT.

Today, I'll share how I went from zero to completing the first version.

1. How Much Can AI Help When Building Games Solo?

Company game roles include Producer, Designer, Programmer, Artist, and Tester.

Programming used to be the highest barrier; now everyone can be a genius programmer thanks to AI.

Design has become the most valuable skill. During development, I found AI cannot replace smooth gameplay logic and creative ideas.

As a programmer, I didn't write a single line of code—I just prompted in Chinese. Art assets were also generated directly by ChatGPT.

Only the game flow, system loops, and user experience required my own thinking.

So AI is powerful—it does things you couldn't do before—but core competitiveness still relies on human imagination.

2. Which Game Engine to Choose?

Common engines:

Engine

Key Differences

When to Consider

Cocos Creator

Scene editing in editor; official WeChat packaging support.

Best for WeChat targets like hidden object, match-3, or casual level-based games. Official Docs

Unity / Tuanjie

Full 2D/3D tools, animation, physics. Tuanjie is Unity China's version with WeChat deployment.

If you have Unity experience or need complex 3D scenes/animations. Check release paths. Engine Intro, WeChat Deploy

Godot

Free, open-source, no royalties. Supports 2D/3D export to desktop/mobile/web.

For indie devs willing to learn scripting who want full control. Official Intro

GDevelop

Visual event-based logic (e.g., click item -> add score). Includes AI assistance.

If you don't want to start with coding and just want to prototype quickly. Intro, Publishing

You can choose based on your preferences. I love all types but especially puzzles, so I started with 2D puzzles.

Based on this, I chose the WeChat Developer Tools 😂. Yes, I didn't use any game engine because I felt it was overkill for this small project.

Of course, you don't need to master these engines. Just ask AI which one fits your game type—it knows everything.

3. The First Level

After choosing the tool, I started building the first level.

My brain couldn't invent a completely new game, so I stood on giants' shoulders.

I found a good existing game, played it, and decided to replicate one level to test the workflow.

I described my playthrough to AI: "Make a simple mechanic where players find all characters in a scene. Most are easy to spot, but some require thinking or checking."

It gave me a basic test prototype:

宋宋 - inline image

It was very crude, but matched my idea.

Once the test passed, I asked it to add style, generate resources, and write code. Final result:

宋宋 - inline image

This is pre-release. Scan the QR code in the title image to try it. I dare say 98% of people won't find all characters without hints.

Here is the systematic prompt I used to initialize the project:

text
1You act as the Technical Lead for this project, responsible for tech research, architecture design, implementation planning, and quality control.
2
3My goal isn't just to make a playable mini-game, but to build a maintainable, scalable WeChat mini-game engineering foundation supporting future different game types using [Target Engine].
4
5Support judgments with source code, official docs, and actual verification. Don't substitute evidence with phrases like "senior experience," "best practices," or "mature frameworks."
6
7I. Project Background & Gameplay Needs
8
9I will provide:
10- Game Idea: [What kind of game]
11- Core Experience: [What players should feel]
12- References: [Games, videos, images, docs, specific parts]
13- Target Platform: [WeChat Mini-Game, Web, Mobile, PC, Console, TBD]
14- Engine/Tech Stack: [Confirmed options or TBD]
15- Current Status: [Idea only, Prototype, Existing Project, Launched]
16- Constraints: [Staff, time, budget, devices, resources]
17- Deliverables: [Design doc, Playable prototype, Feature, Release-ready version]
18
19First game type/mechanics determined by my inputs. Organize ideas into developable, verifiable rules covering:
201. Player Goals & Experience.
212. Core Loop.
223. Operations & State.
234. Session Structure.
245. Feature Scope.
256. MVP Validation.
26
27If reference video provided, break down by actual operation order. Distinguish observed facts, inferred rules, my requirements, and your proposals. Cite sources for judgments. Mark inaccessible/unconfirmed items as pending.
28
29If existing project, verify README, entry points, current implementation, confirmed scope. If none, start from verifiable gameplay/platform. Ask minimal necessary questions if critical info missing. Proceed with reasonable assumptions for reversible details, documenting them.
30
31II. Core Objectives
32
33Foundation must support:
34- Easy level addition;
35- Art replacement/iteration;
36- Interaction capability expansion;
37- New gameplay modules;
38- Reusability for other game types;
39- Reliable cloud data sync;
40- Controlled content updates, publishing, monitoring, rollback.
41
42Cross-game reuse doesn't mean stuffing all games into one puzzle engine. Accounts, resources, saves, UI, platform APIs, publishing, monitoring can be shared; different games should have their own gameplay runners.
43
44III. Research Mature Solutions First
45
46Prioritize:
471. Local existing projects.
482. Official WeChat/cloud docs.
493. Well-maintained GitHub open-source projects.
504. Real source code, issues, tests related to project.
51
52Compare candidates on:
53- Feature coverage;
54- Compatibility with engine/WeChat;
55- License/commercial restrictions;
56- Maintenance status/version;
57- Coupling/intrusiveness;
58- Tests/known issues;
59- Integration/upgrade/exit costs.
60
61Don't rely solely on stars/README claims.
62
63Conclude each candidate: Direct reuse / Adapted reuse / Mechanism reference only / Defer / Exclude.
64
65Avoid assembling overlapping resource/UI/event/lifecycle frameworks.
66
67Independently review tech choices in existing solutions. If implementing custom base modules, justify why existing ones fail and define minimal scope.
68
69IV. Architecture Questions to Answer
70
711. Module Boundaries: Define responsibilities/dependencies for app entry, services, platform adapters, gameplay rules, presentation, content systems, save sync, backend. Avoid giant global managers, circular deps, hardcoded level checks, scattered wx/cloud SDK calls.
72
732. Gameplay Abstraction: Input -> Intent -> Precondition check -> Atomic state change -> Events -> Goal update -> Visuals. Separate fact vs presentation, player goals vs cutscenes, session state vs permanent progress. Prove abstraction with known levels. Specify config-only changes, new resource needs, code additions, new module needs.
74
753. Resources & Lifecycle: Shared refs, ownership, release timing, object pools, async cancellation, late callbacks, event/timer cleanup. Focus on edge cases: switching levels, closing UI, load failures, repeated opens, backgrounding—not just happy paths.
76
774. Art & Content Production: Stable IDs, states/forms, anchors, logical canvas, crop offsets, layers, hit areas, placement zones, color masks. Organize source/runtime/build artifacts. Check missing/misaligned/incompatible assets.
78
795. Database & Sync: Local vs Cloud vs Server-authoritative data. Identity boundaries, models, indexes, API roles, transactions, idempotency, migrations. Cover offline queues, timeouts with server success, duplicate requests/rewards, multi-device conflicts, guest binding/account switch, corrupted saves, old versions, client forgery. "Client uploads whole JSON" is not a complete sync solution. Client cannot decide currency/rights.
80
816. Updates & Dynamic Levels: Distinguish WeChat code package updates, remote resource updates, level data updates, operational config updates per official docs. Explain publishing for new levels/components/algorithms. Chain: Create -> Validate -> Build -> Upload -> Pre-release -> Compat Check -> Gradual Rollout -> Monitor -> Rollback. Consider old clients, unknown capabilities, partial uploads, cache, missing files, ongoing sessions, DB changes. Don't apply native App hot-update logic to WeChat. Technical feasibility != Platform permission.
82
83V. Proactively Fill Engineering Gaps
84
85Check/add:
86- Ad lifecycle/reward deduplication;
87- Foreground/background switching, input cancellation, safe area/device adaptation;
88- First-screen loading, bundle size, texture memory, low-end performance;
89- Network errors, maintenance mode, offline fallback;
90- Environment isolation, key management, permissions, log sanitization;
91- Content backend, publish permissions, audit logs;
92- Monitoring, cost budgets, backup/recovery;
93- Automated tests, version compatibility, fault injection;
94- Third-party code/asset licensing;
95- Foundation upgrades, template maintenance, second game integration.
96
97Classify as "Must-do for MVP / Do later / Skip now" to avoid over-engineering.
98
99VI. Phase 1 Deliverables
100
101Provide reviewable files, not just chat concepts:
1021. Tech Selection & GitHub Reuse Review: Links, versions, licenses, tradeoffs, read-code scope, unverified items.
1032. Engineering Architecture Plan: Module roles, dependencies, directory structure, data flows, key decisions.
1043. Data & Sync Plan: Models, auth, transaction boundaries, retries, idempotency, conflicts, migrations.
1054. Content Production & Update Plan: Asset protocols, level creation flow, version compat, dynamic levels, publish/rollback examples.
1065. Phased Implementation & Acceptance Checklist: Inputs/faults, expected results, validation env, pass criteria.
107
108Validate plan with existing 4-level mechanics, then test reuse boundary with a different gameplay type.
109
110VII. Quality & Execution Requirements
111
112- No zero-bug promises; convert risks to verifiable cases.
113- Don't mix static analysis, issue reports, design reasoning, and runtime results.
114- Mark unexecuted tests as such.
115- Don't treat empty dirs/interfaces/managers as mature frameworks.
116- Don't introduce microservices/ECS/script interpreters/no-code editors unless needed for reuse.
117- Verify one complete chain first: Level creation, play, save, cloud settlement, restore, content publish, rollback.
118- Tech Lead handles automatable engineering; don't dump all ops on non-tech users.
119- Ask minimal questions only when truly necessary; continue otherwise.
120- Phase 1: Research/design only. No modifying formal game project, installing external extensions, activating paid services, deploying, or publishing.
121- After confirmation, implement in phases; deliver runnable results and acceptance evidence per phase.

4. Workflow Changes

I built levels one by one. The first two took days because results weren't right, and I was figuring out resource layering/ordering on the fly. Styles kept changing.

I realized even "simple" games are hard to replicate with AI.

AI lacks precision/detail; requires manual correction. It often tries to fool you with rough versions.

How to improve efficiency/quality?

Some things are learned by doing—experience.

For asset generation, I tried several methods: full image then cutouts (poor quality); separate interactive objects (felt disjointed/inconsistent).

Final solution: Generate logical full images first, then use a specific skill to layer interactive objects, keeping style less realistic. Then iterate endlessly 😂.

Test demo for new level (gameplay only, no assets):

宋宋 - inline image

Still crude, but fresh excitement initially.

When I only made 2 levels in a week, I started looking for shortcuts.

5. Starting the Shortcut Method

The game needs many levels. The reference game had 1300+ levels. My first version needed at least 30 unique levels to keep players engaged.

Previous method was slow and lazy (especially during holidays). Making 10 levels would take forever.

New trick: Record gameplay of another game's level, send video to ChatGPT, say "Replicate this roughly, change style slightly, add/remove few mechanics, give me a playable level."

Surprisingly effective. It quickly built a framework. Based on that, I added ideas/opinions iteratively. Could generate multiple levels simultaneously. My job became giving feedback/modifications.

Got to 50 levels this way, tweaking constantly.

Turned "Video-to-Level" into a skill.

Finally felt the game had potential.

Two main optimizations for game dev:

  1. Show screenshot to AI: "Optimize this, optimize that." 😂
  2. Seriously, good games are polished slowly. Mature games aren't designed instantly; they're refined over time.

6. Believe in Yourself

As a game programmer, my daily work involves adding systems, optimizing, server features, iterating in large games.

Yet I doubted myself when starting solo.

But I love playing games.

So why not?

Started with doubt/confusion.

Thought I had no ideas, but questions naturally popped up.

Solving those questions + max execution = maybe fate turns.

Not fully satisfied yet, but a prototype exists.

Now, in various environments, I automatically think "This scene could be a level." That's passion.

If you like games, check out what an ordinary person built with AI.

Most importantly: Can this make money? I'll update later.

I'm Song Song. Thanks for reading. X: @songsong

Remix in YouMind

Turn one viral article into a full content workflow

Collect the source, decode the pattern, create assets, draft the story, and distribute from one AI workspace.

Explore YouMind
For creators

Turn your Markdown into a clean 𝕏 article

When you publish your own long-form writing, images, tables, and code blocks make 𝕏 formatting painful. YouMind turns a full Markdown draft into a clean, ready-to-post 𝕏 article.

Try Markdown to 𝕏

More patterns to decode

Recent viral articles

Explore more viral articles