How Kuikly’s AI Engineering Boosted Cross‑Platform Development Efficiency

Introduction

AI is rapidly transforming software development, and AI engineering has become a priority for many product teams. The Sogou Input Method team shares their practical experience of applying AI coding within the Kuikly cross‑platform project, aiming to provide reference ideas for other teams.

Background

Kuikly is a widely used cross‑platform framework at Tencent, supporting Android, iOS, HarmonyOS, Web, mini‑programs, and macOS. It powers over 20 products with more than 1,000 pages and daily active users exceeding 500 million. The team began experimenting with AI assistance in late 2025, finding that while AI can generate runnable code, the output often falls short of being truly usable for complete page requirements.

Challenges

Engineering Understanding Gaps : The AI model frequently hallucinates API calls and struggles with existing complex codebases.

Unstructured Requirements : Vibe Coding relies on developers to fill in ambiguous details, leading to inaccurate code generation and high rework costs.

AI‑Friendly Engineering

Improving the underlying codebase quality proved more effective than tweaking prompts. Kuikly’s architecture already separates UI from business logic, maintains clear module boundaries, and provides unified system capabilities, allowing AI to reference existing modules without extensive additional processes.

Precise AI Context

To give AI the right information, the team built an AI Context Document Generator (Skills) that extracts module responsibilities, core APIs, parameter meanings, and dependency relationships into structured documents. These documents are iteratively refined using a GAN‑inspired workflow: a generator creates the doc, an evaluator scores it, and the generator revises until quality thresholds (e.g., ≥95 % in the fifth round) are met.

Standardized Requirement Process

The team introduced Spec‑Kit, a three‑stage pipeline ( /speckit.spec → /speckit.plan → /speckit.tasks) that converts product requirements, design drafts, and interaction specs into structured engineering artifacts. AI drives most of the transformation, with human checkpoints at key milestones.

Kuikly AI Toolset

Kuikly provides a suite of AI‑assisted capabilities, including MCP services, Rules, Skills, AI debugging, D2C code conversion, and AI transcoding. These tools cover the entire development lifecycle—from code generation to debugging and migration. The CLI allows one‑click updates to keep the AI capabilities in sync with framework changes.

Practice and Results

Using the new workflow, the team built the "Inspiration Word Library" page for the latest Sogou Input Method release. The process involved:

Generating a complete Spec‑Kit documentation set.

AI‑driven code generation based on the spec.

Minimal UI adjustments and no major architectural refactoring.

Compared to traditional development (≈3 days of pure coding and technical design), the AI‑enhanced flow delivered the same functionality in 1 day. Measured improvements include:

Documentation accuracy matching source code.

100 % API coverage.

Example code ready for copy‑paste.

Onboarding speed ↑ 50 %.

Module comprehension speed ↑ 107 %.

AI coding accuracy ↑ 114 %.

Documentation update efficiency ↑ 137 %.

The gains were most pronounced for new, well‑scoped pages; legacy code modifications still require manual intervention, and stability varies across scenarios.

Outlook

Future work focuses on:

Integrating D2C tools to automatically convert design drafts into high‑quality code, reducing UI‑fix effort.

Building automated verification mechanisms that use AI to validate UI rendering, interaction behavior, and layout consistency.

Expanding AI assistance to tasks such as requirement parsing, bug localization, cross‑module refactoring, and end‑to‑end delivery pipelines.

About Kuikly

Kuikly is open‑source; interested teams can access the repository and documentation via the links below.