Kuikly AI: Harness Engineering Foundation for Cross-Platform Dev, Rules & Skills Open-Sourced

Background: AI Reshaping Software Engineering

Over the past two years, AI has profoundly reshaped software engineering. Developers increasingly integrate AI into daily workflows, achieving significant coding efficiency gains. Two paradigms have emerged: Vibe Coding — using prompts to let large models generate code directly; and Spec Coding — guiding AI with structured requirements and technical specifications for more controllable, expected output. Engineering methodology has evolved from Prompt Engineering (crafting prompts) to Context Engineering (managing context), and now to Harness Engineering — building a complete constraint system that makes AI development controllable, verifiable, and stably deliverable.

Harness Engineering proposes a core idea: LLMs need an engineered "rein" — a control system around the LLM/Agent that constrains, validates, and continuously manages uncertain model outputs within deterministic processes. This system typically comprises five aspects:

System Prompt & Context Injection

Tools & Execution Environment

Control & Orchestration Logic

Persistence & State Management

Observe & Verify

Challenges in Client-Side AI Adoption

When AI coding tools enter real client-side development, new challenges appear:

Uneven generated code quality : AI output is voluminous and runnable but often "wild" — non-standard, inelegant, with wrong dependencies, non-existent API calls, messy structure, and framework-noncompliant patterns.

Inherent client-side complexity : Clients target multiple runtime environments; projects often coexist with multiple architectures and tech stacks, carrying complex legacy debt and component dependencies. Design-to-code requires multiple visual-fidelity rounds, while client runtime verification chains are long and slow — exactly where current AI struggles most.

These issues hinder AI capabilities from reaching engineered deployment and scaled application in client scenarios.

Kuikly AI: A Full-Stack AI Engineering Capability System

Kuikly is a lightweight, high-performance, dynamically-updatable cross-platform framework widely adopted inside and outside Tencent. In the AI wave, Kuikly explores an AI-native path: not just enabling AI to write Kuikly code, but to write good Kuikly code. The Kuikly AI project builds a complete engineering capability system covering the entire client development workflow, systematically injecting framework knowledge, component specifications, best practices, and toolchains into AI workflows. The framework bears the AI adaptation burden so users focus on business and get standardized, high-quality, out-of-the-box Kuikly development experiences.

The system includes eight core capabilities:

AI Q&A Agent : Knowledge base, guidance docs, and code RAG to improve answer accuracy/depth, lowering onboarding barriers and resolving common usage issues.

Kuikly Rules : Injects context into models — development specs, architecture principles, best practices — compensating for LLM's lack of Kuikly understanding and improving code standard compliance.

Kuikly Skills : Scenario-specific skill packs enabling AI to handle complex tasks like an expert, achieving high-quality intelligent code generation.

Kuikly MCP : Based on Model Context Protocol, lets AI tools dynamically access official Kuikly docs, component library indexes, and dev tools for real-time knowledge, eliminating stale information.

Visual-to-Code (Deco) : Converts Figma designs to Kuikly code automatically, precisely and efficiently restoring design intent, reducing page scaffolding effort.

Preview & UI Inspector : Instant preview of AI-generated code runtime effects, with UI Inspector for visual component tree and property inspection, closing the critical "Observe & Verify" gap.

Transpilation Agent : Migrates existing React/Vue/Hippy code to Kuikly efficiently, auto-analyzing structure and component relationships, mapping syntax and APIs for smooth migration.

AI Chat Component : Ready-to-use official AI chat component for in-app AI dialogue, supporting multi-model switching, cross-platform consistent UX, and highly customizable UI.

Development Agent : Orchestrates the full workflow from requirement planning, code generation, multi-platform preview to one-click verification, integrating scattered AI capabilities into a smooth automated production line so developers focus on business logic.

First Open-Source Batch: Rules & Skills

3.1 Rules

KuiklyDSL is a proprietary DSL; current LLM training corpora lack sufficient coverage, so without rule constraints models easily generate unusable code. Kuikly ComposeDSL, while based on the Compose paradigm, differs from official Jetpack Compose in package imports, component APIs, and parameter definitions, causing models to misuse them. Therefore, targeting high-frequency issues in both DSLs, corresponding Rules are provided:

KuiklyDSL Rules : Cover component usage, API specifications, parameter constraints, and common error patterns.

ComposeDSL Rules : Address package import differences, API mapping, parameter definition conventions, and migration pitfalls.

Beyond official general rules, adopters can further supplement their own project architecture conventions, directory standards, and best practices, letting AI work "in your engineering, your way."

3.2 Skills

Kuikly Skills serve as a skill library for AI Agents (CodeBuddy, Cursor, Claude, etc.), giving AI clear specifications and guidance when facing complex Kuikly development processes, enabling more efficient task completion. Currently, 11 distinct skills cover all steps of Kuikly development: framework integration, UI construction, basic development capabilities, and capability extension.

As the skill system matures, AI's role in Kuikly projects will expand beyond code generation to become an engineering assistant participating in the full development lifecycle. Future investments will include performance optimization and component marketplace capabilities.

3.3 Installation & Usage

Installation is simple via terminal:

// Install
npx skills add Tencent-TDS/KuiklyUI-AI/skills

// Update
npx skills update Tencent-TDS/KuiklyUI-AI/skills

Usage examples:

"Create a Todo list page with Kuikly"
"Add an entrance animation to view xx"

3.4 Adoption Feedback & Recommendation

From business lines already onboarded, Rules and Skills effectively enhance model understanding of the Kuikly framework across scenarios: iterating new features on existing projects, building common page layouts, or creating new Kuikly pages from scratch. Developers report significantly improved code standard compliance and usability, smoother overall process, and many framework-detail issues that previously required repeated fixes are now preempted at generation time — often only minor style or interaction tweaks remain. The team recommends using both Rules and Skills together for the best experience.

Roadmap

Rules and Skills are just the first step. Following the "make AI understand client-side better" mainline, the team will continuously release more engineered capabilities, prioritizing MCP service, visual-to-code, preview & inspector, transpilation agent, AI chat component, and development agent — rolling them out progressively.

References

Kuikly is now open source. Repository: https://github.com/Tencent-TDS/KuiklyUI. Official documentation: https://kuikly.tds.qq.com/Introduction/arch.html. Kuikly belongs to the Tencent Terminal Service Alliance (tds.qq.com). More information at https://tds.qq.com/ and TDS Framework site https://framework.tds.qq.com/.