3 Hours, Zero Code: Building a WeChat Mini Program with AI Agents in Practice

Starting from a Real Need

The author's child is in second grade. The teacher requires parents to record videos of their children reading textbooks aloud every day and send them via WeChat. With over fifty students in the class, the teacher's WeChat is flooded with check-in videos daily. Posting to the group chat? That would disturb other parents.

This pain point sparked an idea: Build a WeChat Mini Program where parents upload their check-in videos individually, and the teacher can review them in batches at any time — without disturbing anyone.

Abstracting this requirement further — it's essentially a "public check-in activity" system, applicable to any team, organization, or group's daily check-in scenarios.

The author dictated the requirements, organized them into text using a note-taking tool, then pasted them directly into a self-built Mini Program development framework. Throughout the entire process, actual human involvement was only about 30 minutes — the remaining two-plus hours were handled entirely by AI Agents running autonomously.

Core Method: Using AI Workflows to Output Code

The author offers an insightful analogy:

In traditional software development, we use code to process data; in AI Agent development systems, we use AI workflows to output code.

The core idea of this framework is: orchestrate the complete software development process (requirements → design → development → testing) into an automated Agent pipeline, with each stage handled by AI Agents playing different roles.

AI Agents are AI systems capable of perceiving their environment, making autonomous decisions, and executing actions to achieve goals — fundamentally different from traditional script-based automation. Traditional automation relies on predefined rules and fixed processes, failing when encountering unexpected situations; AI Agents can understand context, handle ambiguous instructions, and dynamically adjust strategies. In software development scenarios, Agents can not only write code but also understand requirement intent, identify potential issues, and autonomously decide next steps. Multi-Agent Systems further decompose complex tasks among Agents with different specializations, similar to a virtual development team where each member has their own role, collaborating through structured information exchange to accomplish the overall objective.

Process Design: Multi-Version Divergence + Fusion

The entire zero-code development process is divided into the following key stages:

1. Raw Requirements → Input into requirements document
2. Product Manager (Multi-Version Divergence) → Leverage model randomness to generate multiple product design proposals, then fuse them into a single PRD
3. Prototype Designer (Multi-Version Divergence) → Similarly generate multiple interaction prototype versions, cross-pollinating strengths, then fusing into the final prototype
4. UI Review → Visual review of the prototype
5. Architect → Design technical architecture (also using multi-version divergence)
6. Development → Output complete Mini Program code
7. UI Review + Fixes → Specialized review for WeChat Mini Program UI
8. Testing + Iteration → Discover issues and auto-fix

PRD (Product Requirements Document) is a core document in software engineering that bridges business requirements and technical implementation, typically containing feature descriptions, user stories, interaction logic, boundary conditions, and acceptance criteria. In traditional development workflows, PRDs take product managers days or even weeks to write, serving as the project's "constitution." Having AI play the product manager role to automatically generate PRDs means the AI needs advanced capabilities including requirements analysis, user scenario modeling, and feature priority assessment. Notably, the quality of AI-generated PRDs directly determines the output quality of all subsequent stages — which is precisely why the author adopts a "multi-version divergence + fusion" strategy, using multiple generation rounds to improve PRD coverage and soundness.

Why "multi-version divergence"? Large language model outputs are controlled by the Temperature parameter — higher temperature yields more random and diverse outputs; lower temperature yields more deterministic and conservative outputs. The "multi-version divergence + fusion" strategy is essentially an application of Ensemble Learning principles in generative AI: obtaining multiple candidate solutions through multiple samplings, then extracting quality features from each through a fusion step, ultimately producing results superior to any single output. This mirrors "code review" and "design review" mechanisms in software engineering — the collision of multiple perspectives often reveals blind spots invisible from a single vantage point, and AI can complete in minutes the multi-round review iterations that would take human teams days. This is far more pragmatic than trying to get the model to output a perfect result on the first try.

Framework Structure Deep Dive

Agent Definitions

All Agents are defined in the agents directory, each with a standard structure:

• Name and Description: Clear role positioning
• Available Tools: Defining the capabilities an Agent can invoke
• Work Standards and Artifacts: Universal specifications independent of specific business logic

These definitions are reusable — swap in a different requirements document, and the same framework can develop another WeChat Mini Program. This means you build the framework once and reuse it across different projects.

One-Click Automated Development Launch

The author defined a custom command allbegon — after filling in the requirements, executing this single command kicks off the entire Agent workflow automatically.

On the tech stack side, the tool used is Claude Code — Anthropic's command-line AI programming tool for developers, built on the Claude model, specifically optimized for code generation, debugging, and engineering tasks. Unlike conversational AI, Claude Code can directly read and write the local file system, execute terminal commands, and manage project directory structures — possessing genuine "hands-on capability." Its core advantage lies in its ultra-long context window (supporting hundreds of thousands of tokens), enabling it to understand the context of an entire codebase simultaneously rather than processing individual files in isolation. By adding specific parameters to prevent the AI from interrupting or asking questions during execution, long-duration autonomous operation is achieved.

The model selected is Zhipu (which performs well among domestic options), with concurrency controls implemented to address its rate-limiting issues. Domestic LLM APIs generally employ Rate Limiting mechanisms, typically measured in RPM (requests per minute) and TPM (tokens per minute). In multi-Agent concurrent execution scenarios, if large numbers of API requests are fired simultaneously without control, rate limits are triggered causing task failures. Common mitigation strategies include token bucket algorithms for request rate control, exponential backoff retry for handling failed requests, and request queuing for managing concurrency caps — these are critical engineering details for taking AI tools from "lab-usable" to "production-ready."

Once launched, the AI automatically completes the following:

• Creates the project directory structure
• Generates 5 product versions with different characteristics
• Progressively completes the entire design, development, and testing workflow

Throughout this process, the author can go do other things, and when execution completes, the result is a usable WeChat Mini Program product. If minor issues exist, they can be fine-tuned through conversation at the end.

It's worth noting that the technical characteristics of WeChat Mini Programs themselves are an important prerequisite for this approach's successful implementation.