How AI Coding Transformed Kuikly’s Cross‑Platform Development: A Practical Playbook

Background

In the AI era, many product teams aim to achieve end‑to‑end AI‑assisted coding, hoping to automatically link requirements, generate code, and test results. The Sogou Input Method team, using the Kuikly cross‑platform framework (supporting Android, iOS, HarmonyOS, Web, mini‑programs, and macOS), sought to push AI coding beyond simple demos to real engineering productivity gains.

Challenges identified included the framework’s “hallucination” of APIs, the tendency to reinvent wheels, and the difficulty of handling unstructured requirements that lead to multiple revision cycles.

AI Engineering Approach

The team first experimented with Vibe Coding, then integrated Kuikly’s own AI tools for component development and small‑scale tasks. They defined a set of key layers for moving from speculative coding to a fully engineered workflow, termed Spec‑Kit.

Building Precise AI Context

To avoid overwhelming the model with irrelevant code, the team built an AI context generator that extracts structured documentation for each module, including module responsibilities, core APIs, parameter meanings, and dependency relationships. This documentation is iteratively refined using a GAN‑inspired multi‑round generation‑evaluation loop, with quality thresholds increasing from 75% to 95% across five rounds.

The resulting context documents achieved 100% source‑code accuracy, full API coverage, and ready‑to‑copy example code, leading to a 50% faster onboarding, 107% faster module‑function understanding, 114% higher AI coding accuracy, and 137% faster documentation maintenance.

Standardized Requirement Process

Without a standardized process, AI‑human collaboration is ad‑hoc and hard to reproduce. The team introduced a Spec‑Kit pipeline consisting of three stages: /speckit.spec → /speckit.plan → /speckit.tasks. This transforms product requirements, design drafts, and interaction specs into structured engineering documents, with AI leading the process and human checkpoints at key points.

The pipeline enforces consistent quality, captures knowledge for reuse, and shifts AI collaboration from improvisational prompting to specification‑driven execution.

Kuikly’s Out‑of‑the‑Box AI Tools

Kuikly provides a suite of AI capabilities—MCP service, Rules, Skills—as well as debugging, D2C, and AI transcode tools. By leveraging Rules, Skills, and MCP, the team reduced AI hallucinations and ensured the model understood framework constraints, allowing developers to focus on business logic rather than model training.

A CLI tool enables one‑click updates of AI capabilities, keeping the framework and AI tools synchronized without manual intervention.

Practice and Results

Using the new AI‑driven workflow, the team built the “Inspiration Word Library” page for Sogou Input Method. The page required dynamic multi‑column layouts, dark‑mode switching, network requests, routing, KV storage, and analytics.

Through Spec‑Kit, the requirements were transformed into a complete engineering document, and AI generated a high‑fidelity implementation in one day—compared to three days of traditional development. The final output consisted of 16 Kotlin files, roughly 2,000 lines of code, covering card lists, dynamic layouts, skeleton screens, error/empty states, and dark mode.

For new modules and pages, the AI pipeline consistently reduced development time and maintained architectural standards, though legacy code modification and deep refactoring still require manual oversight.

Future Outlook

The team plans to integrate D2C tools for automatic UI code generation, automate verification of UI rendering and interaction consistency, and extend AI assistance to requirement parsing, bug fixing, and cross‑platform coordination.

About Kuikly

Kuikly is an open‑source cross‑platform framework maintained by Tencent’s TDS. Interested developers can explore the GitHub repository and official documentation for further details.