PomLink: Modular iOS Agent for LLMs

• PomLink is an iOS agent prototype using POML to modularize prompt engineering and integrate documents, tables, images, and audio into LLM-powered conversations.
• It employs an HTML-like markup with CSS-inspired styling and external JSON directives to decouple content logic from presentation, enhancing maintainability.
• Empirical evaluations reveal that PomLink’s component-based design improves LLM accuracy and developer productivity by reducing error rates and streamlining prompt modifications.

PomLink is an iOS agent prototype built using the Prompt Orchestration Markup Language (POML), designed to demonstrate how prompt engineering for LLMs can be modularized, structured, and efficiently integrated with complex, multimodal data within a conversational agent framework. By leveraging POML's component-based markup, PomLink enables seamless “linking” of various file types—including documents, tables, images, and even audio memos—into LLM-driven conversations, thus offering a maintainable and extensible architecture for sophisticated LLM applications.

1. Design Philosophy and Framework

PomLink’s design centers on the modular and declarative principles of POML, employing HTML-like tags to encode logical prompt structure. Each functional region of the agent’s prompt—such as the definition of roles, tasks, system instructions, and few-shot examples—is encapsulated in self-contained markup elements (e.g., <role>, <task>, <include>, <conversation>). This composable approach facilitates both reusability and clarity, ensuring that modifications in system behavior, user experience, or integration patterns can be localized to specific markup components rather than diffused across ad hoc code. The prototype achieved rapid completion—within two days—by leveraging POML’s rich built-in data and formatting primitives.

The inclusion of a CSS-inspired styling mechanism further decouples prompt logic from presentation concerns. Global style directives can be specified in external JSON files, allowing uniform adjustments to tables, captions, or other textual elements across different prompt contexts without changing the markup itself. This supports not only maintainability but also systematic experimentation with prompt format variations for optimized LLM response accuracy.

2. Application and Data Integration

PomLink operationalizes complex data integration scenarios via POML's specialized tags. The agent can draw on external resources using components such as <document> and <table>, allowing, for example, PDF sections or spreadsheet data to be directly embedded in the conversational prompt. The markup also enables straightforward inclusion of web content and voice transcripts. The <include> tag facilitates compositional prompts by importing pre-defined system instructions or reusable boilerplate segments.

This architecture supports advanced agent workflows: PomLink can ground its responses in user-uploaded files, perform contextual cross-referencing, and maintain stateful conversations involving disparate data types. Data inclusion is managed at the markup level, abstracting the complexity of input serialization, context window budgeting, and format normalization away from the agent developer.

3. Impact on LLM Accuracy and Usability

Although the TableQA case paper provides detailed numerical results on prompt styling effects, PomLink serves as a practical demonstration that componentized and styled prompts indirectly benefit both LLM accuracy and developer productivity. Through the compartmentalization of prompt fragments and systematic grounding of data, PomLink reduces error rates caused by brittle, hand-tuned text assembly, and allows for the rapid iteration and swapping of prompt variants. In practice, the use of data-driven style constraints—e.g., for table layout informed by TableQA “Auto” format findings—enables PomLink to pass optimally structured contexts to the LLM backend, supporting improved factuality and consistency in outputs.

4. User Study Evaluation

A user paper evaluating PomLink within the broader POML ecosystem gathered empirical feedback from researchers and developers performing analogous data integration tasks. Participants highlighted several advantages:

Aspect	Feedback	Impact
Data linking / inclusion	Markup enabled rapid file connection in prompts	Reduced manual formatting burden
Content/presentation split	Clear separation via stylesheets improved maintainability	Faster prompt modifications, less ambiguity
Tooling	Live previews, inline error checking in VS Code extension	Streamlined debugging, increased development speed

These findings underscore that structured markup and robust tooling substantially decrease prompt engineering overhead, foster maintainable workflows, and lower the barrier for building multimodal LLM agents like PomLink. Some suggestions for improvement focused on further simplification of core tags and expanded documentation of advanced features.

5. Technical Implementation

PomLink is implemented atop POML’s core technical stack, which comprises:

• Component-based markup language: An HTML-like syntax lets developers specify logical prompt structure and data inclusion points.
• Templating engine: Support for variable substitution ({...}), control flow (e.g., for attributes), and conditional rendering enables dynamic prompt generation based on runtime data.
• CSS-like styling system: Style directives are provided externally (JSON) and control presentation aspects independently of content logic.
• Three-pass renderer: The markup undergoes parsing and validation (first pass), transformation to an intermediate representation (with resolved style and metadata; second pass via React), and serialization to an LLM-ready format (third pass) such as Markdown or plain text.
• Developer toolkit: Integration with modern IDEs (notably VS Code) provides live render previews, inline diagnostics, and auto-completion, enhancing the end-to-end application development lifecycle.
• Backend SDK: Node.js-based SDK processes dynamic data binding and connects the markup-rendered output to the actual LLM service.

Each of these layers contributes to system robustness, facilitating iterative prompt refinement and the scalable development of agentic applications.

6. Future Directions and Implications

PomLink’s deployment highlights several avenues for further research and engineering:

• Expansion of modal coverage: As LLMs gain increased multi-modal capability, PomLink (via POML) can adopt new components for video, interactive graphs, or other advanced data types.
• Automation of prompt optimization: The generality of the content/style separation allows for future systems to autotune prompt layouts (as demonstrated in TableQA), potentially in response to real-time performance signals.
• Broader platform integration: Experience from iOS prototype development indicates viability for parallel support on Android or web clients.
• Enhanced tooling: Refinement of developer-facing features (e.g., clearer diagnostics, better accessibility, richer documentation) is anticipated to further lower the technical threshold for production deployment.
• Generalization to agent frameworks: The PomLink experience provides an archetype for LLM-based agents that compose structured prompts, integrate heterogeneous data, and maintain conversation state, laying the foundation for future frameworks targeting robust cross-source knowledge integration in conversational AI.

7. Contextual Significance

PomLink directly exemplifies POML’s capacity to address prompt engineering challenges endemic to LLM applications: structure, formatting sensitivity, integration complexity, and developer productivity (Zhang et al., 19 Aug 2025). Its architecture and empirical demonstration make it a noteworthy reference for researchers developing scalable, maintainable, and multimodal LLM-powered agents. The broader implication is a methodological shift toward declarative, component-based prompt design for LLM applications—an approach evidenced to scale not only development workflows but also the performance and reliability of conversational AI systems.