跪拜 Guibai
← All articles
AI Coding · VibeCoding · Android

A Full AI-Driven Software Workflow, from PRD to Review, Without Writing a Line of Code

By hqk ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

This workflow shifts AI coding from one-shot prompts to a repeatable, team-like process. The key insight — that project documentation, not chat history, is the real long-term memory — means any new AI session can pick up exactly where the last one left off, cutting token waste and inconsistent decisions.

Summary

Starting from an open API, a developer ran a complete software R&D workflow using only AI, never writing code. The AI first analyzed the API and produced a PRD, roadmap, and architecture docs. When Figma's MCP hit a paywall, the AI built its own JSON-driven wireframe viewer to keep the prototype phase free and machine-readable. A design system was then translated into Compose runtime mappings — Theme, Component, and Page mappings — so the AI could generate UI code directly from specs. Before any feature work, an engineering bootstrap laid down shared infrastructure: dependency injection, navigation, theme, error handling, and platform contracts. A "Ponytail" skill was installed to enforce minimal implementation, preventing the AI from over-engineering. The project's AGENTS.md was refactored into a feature development guide that freezes design assets, defines a source of truth, and sets a Definition of Done. With that foundation, a single prompt — "Develop the homepage" — triggered a full plan, impact analysis, coding, build verification, and review cycle. The result is a runnable Kotlin Multiplatform app and a reusable workflow where the prompt shrinks as the project's memory grows.

Takeaways
A complete KMP app was built without writing code by assigning AI the roles of PM, UX, visual designer, engineering lead, and reviewer.
AI analyzed the WanAndroid API docs and produced a PRD, roadmap, and architecture before any coding began.
Project memory was stored in repo documents (PRD, RoadMap, Architecture, Task/Bug Memory, ChangeLog), not chat history, so new sessions restore state instantly.
When Figma's MCP hit a free-tier limit, the AI built a custom JSON-driven wireframe viewer (AIWireframeViewer) to replace it.
A Design System was translated into Compose Runtime Mappings (Theme, Component, Page) so AI could generate UI code directly from specs.
An Engineering Bootstrap phase built shared infrastructure — DI, navigation, theme, error handling, platform contracts — before any feature pages.
The Ponytail skill was installed to enforce minimal implementation: only add new code when no existing capability, platform API, or dependency already solves it.
AGENTS.md was refactored into a Feature Development Guide with Frozen Assets, Source of Truth, Ponytail Policy, and a Definition of Done.
A single prompt — "Develop the homepage" — triggered a full plan, impact analysis, coding, build verification, and review cycle.
Review is not a separate phase; the AI performs build checks, code review, and memory updates as part of every feature completion.
Conclusions

Prompt length shrinks as project knowledge grows. The endpoint of prompt engineering is project engineering — embedding rules, specs, and memory into the repo so the AI needs only a task name.

AI over-engineers by default. A skill like Ponytail that forces a checklist — can I reuse, can the platform do it, does a dependency already cover it — is essential to keep output minimal and maintainable.

Figma and similar tools are overkill for AI-driven prototyping. A JSON-driven wireframe viewer gives the AI a stable intermediate language that feeds directly into code generation.

Design tokens alone are not enough for AI coding. A runtime mapping layer (Theme Mapping, Component Mapping, Page Mapping) is needed to translate design language into framework-specific code.

Multi-agent architectures are the logical next step: separate agents for product, UX, visual, architecture, and engineering, each with their own memory, boundaries, and Definition of Done, collaborating through documents and rules.

Concepts & terms
KMP (Kotlin Multiplatform)
A Kotlin technology that allows sharing business logic across Android, iOS, and other platforms while keeping platform-specific UI code separate.
Engineering Bootstrap
A Phase 0 step that sets up shared project infrastructure — dependency injection, navigation, theme, error handling, platform contracts — before any feature pages are built, so every subsequent feature inherits a consistent foundation.
Ponytail Skill
A Codex plugin that enforces minimal implementation by checking whether a new capability is truly needed, or if existing code, platform APIs, or dependencies already solve the problem.
Frozen Assets
Product, UX, and visual design documents marked as immutable in AGENTS.md so that ordinary feature coding tasks cannot modify them, preserving design decisions across development.
Design Runtime Mapping
A translation layer that converts design language (colors, fonts, components) into framework-specific code mappings (Theme Mapping, Component Mapping, Page Mapping) so AI can generate UI code directly.
Computer Use
An AI agent capability that simulates real human mouse and keyboard operations on a graphical interface, as opposed to calling APIs directly.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗