Artificial Intelligence

A No-Code Health Agent on Huawei's Xiaoyi Platform: Prompts, Compliance, and What Actually Worked

By 一只牛博 · Jul 15, 2026

Read original on juejin.cn ↗ Google Translate ↗ Alt translation

A no-code agent builder that ships with a capable LLM, real-time debugging, and built-in compliance checklists lowers the barrier for individual developers to publish AI assistants on a major mobile ecosystem. The platform handles API keys, hosting, and regulatory declarations, removing the infrastructure work that usually blocks solo builders.

Summary

A health assistant agent covering BMI calculation, meal analysis, and dietary advice was built from scratch on Huawei's Xiaoyi Open Platform without writing a single line of code. The entire process—selecting a development mode, writing role instructions, and testing—happened inside the platform's web console. The built-in DeepSeek-V3 model handled weight analysis and generated a seven-day meal plan from natural-language prompts, while the debug panel provided real-time chat previews and text-to-speech output. Plugins, knowledge bases, and external tool integrations were deliberately left unused to test how far pure prompt engineering could go. The platform also enforced a content compliance declaration referencing China's generative AI regulations before the agent could be published. Two of the three planned capabilities were verified; meal analysis and boundary-refusal behavior remain untested.

Takeaways

— An agent was built entirely through a web console, with zero custom code or local services.

— The platform's built-in DeepSeek-V3 model required no API key configuration.

— Role instructions included explicit capability boundaries and refusal rules, not just feature lists.

— Real-time debug previews rendered chat bubbles and follow-up suggestions as the prompt was written.

— Content compliance required declaring AI-generated content and any third-party model usage, citing three specific Chinese regulations.

— Long-term memory was enabled by default, which is critical for health agents that need to remember user metrics across sessions.

— Only two of three planned capabilities (BMI analysis and diet planning) were tested; meal analysis and medical boundary refusal were not verified.

Conclusions

Platforms that bundle model access, debugging, and compliance into a single web console remove the three biggest friction points for solo developers: API key management, UI prototyping, and regulatory paperwork.

Writing capability boundaries into the prompt with the same detail as the feature list produced responses that went beyond templated answers—the BMI reply added a 'close to overweight' warning unprompted.

The description field doubles as a search keyword index for the agent marketplace, a detail easy to overlook that directly affects discoverability.

Default-on long-term memory is a quiet but critical design choice for any agent that handles personal metrics across conversations.

Leaving plugins and knowledge bases empty was a deliberate constraint that exposed the ceiling of prompt-only agents: they can handle structured Q&A well but can't access real user data or external services without further configuration.

Concepts & terms

Agent vs. Skill

On the Xiaoyi platform, an Agent is the conversational entity that interacts with users, while a Skill is a reusable functional module (like a快递 query or fitness coach) that an Agent can call.

Cloud Workflow mode

A development mode for orchestrating agents with clearly defined business steps, as opposed to device-local or third-party-model-driven modes.

Content Compliance Declaration

A mandatory step before publishing that requires developers to declare whether their agent contains AI-generated files or accesses external large models, referencing China's generative AI regulations.

Source: juejin.cn ↗ Google Translate ↗ Backup ↗