Wordflow: Social Prompt Engineering

• Wordflow is an open-source, browser-based text editor that enables social prompt engineering by integrating local LLM inference, prompt sharing, and a community hub.
• It lowers entry barriers by embedding familiar text editing workflows with on-device LLM execution for privacy-preserving, accessible prompt development.
• The platform leverages modern web technologies, including Web Components and Lit Element, to support collaborative creation, adaptation, and sharing of prompt templates.

Wordflow is an open-source, browser-based text editor for collaborative prompt engineering with LLMs. It introduces the paradigm of "social prompt engineering," positioning prompt design and reuse as a collective, community-guided process rather than an expert-driven, isolated practice. Wordflow combines familiar word processing workflows with a modular architecture that integrates client-side LLM inference, prompt sharing, and a social hub for prompt discovery. By leveraging modern web technologies, it enables everyday users to create, execute, share, and refine LLM prompts within a collaborative ecosystem, with explicit support for local and privacy-preserving LLM execution (Wang et al., 2024).

1. Definition and Paradigm Shift: Social Prompt Engineering

Wordflow is designed as a text editor with seamless LLM integration, enabling “one-click” application of prompt templates to selected text. The defining innovation is its support for social prompt engineering—a paradigm in which prompts are treated as community artifacts, available for authoring, sharing, curating, and iterative refinement by all users rather than a narrow set of experts.

In this model, prompt engineering is no longer a solitary or ad hoc process but instead becomes an open, cyclical workflow where users:

• Author prompt templates for various tasks,
• Run these prompts to transform text segments,
• Share effective prompts to a public hub,
• Browse and adapt prompts authored by the community, and
• Benefit from aggregated reputational signals (e.g., usage counts, tags).

This approach leverages the collective intelligence and practical experiences of a distributed user base, diffusing prompt engineering expertise to non-expert populations and enabling more robust, accessible LLM applications (Wang et al., 2024).

2. Design Motivations and User Challenges

Prompt engineering presents unstable dynamics and user-access barriers. Small changes in prompt phrasing or structure can lead to drastically different LLM outputs—the so-called brittleness of LLM prompting. Non-expert users typically lack robust heuristics for constructing reliable prompts and are underserved by existing prompt engineering tools, which primarily target AI developers and focus on integration rather than usability.

Wordflow seeks to address these barriers by:

• Lowering entry thresholds for prompt engineering by embedding it in familiar text editing environments,
• Facilitating social learning and the accumulation of best practices through openly shared and discoverable prompt artifacts,
• Enabling on-device execution using local LLMs for privacy preservation, ensuring sensitive text is never transmitted externally.

These motivations tie directly into the system’s architecture and collaborative design, aiming to transform isolated trial-and-error prompt design into a distributed, socially enhanced process accessible to general users (Wang et al., 2024).

3. System Architecture and Enabling Technologies

Wordflow is implemented as a Progressive Web App (PWA) using Web Components and Lit Element, supporting both desktop and mobile usage. Its user interface features four principal modes:

• Editor View: The main interface with a floating toolbar for text selection and prompt application.
• Personal Prompt Library: Stores locally managed prompts.
• Community Prompt Hub: Enables browsing, searching, and importing community-shared prompts.
• Settings Panel: Model selection and configuration.

LLM execution is handled via two primary pathways:

• Client-Side Inference: WebGPU with MLC-LLM enables local execution of open-source models (e.g., Llama 2, Phi-2, TinyLlama)—privacy-preserving and responsive.
• Remote API Access: For commercial models (e.g., GPT-4, Gemini), wordflow can route requests to vendor APIs.

Prompt sharing is mediated by a lightweight backend stack—Amazon API Gateway and DynamoDB—to store prompt metadata, tags, user descriptions, and aggregate run counts.

For text difference visualization, Wordflow employs Myers’s $O(ND)$ diff-match-patch algorithm, highlighting LLM-induced insertions, deletions, and replacements for user review. Local storage is managed through browser IndexedDB, supporting offline and private operation (Wang et al., 2024).

4. Collaborative Prompt Lifecycle: Creation, Sharing, and Adaptation

The collaborative workflow in Wordflow is underpinned by two interrelated libraries:

• Personal Prompt Library: Stores each user’s prompt cards (including title, emoji, template text, and advanced settings like temperature or regex-based output parsing). Prompts may be quickly inserted into the floating toolbar for rapid reuse.
• Community Prompt Hub: Aggregates prompts shared by all users, viewable with recency and popularity sorting and tag filters. Prompts are presented in a detailed viewer (including recommended models and full templates) with one-click import for personal adaptation.

Prompts circulate via an author–share–discover–adapt–reshare cycle. Social signals such as run counts and tag badges guide users towards high-quality or popular prompts, fostering collective filtering and incremental improvement. Although explicit prompt version graphs are not implemented, the evolving community corpus can be conceptualized as a directed acyclic graph where nodes represent prompt versions and edges denote adaptation chains (Wang et al., 2024).

5. Representative Use Cases

Use-case scenarios from the source highlight practical adoption by lay users:

• Technical Writing: A developer discovers and adapts a shared prompt for explaining complex API concepts, applies it with local inference for privacy, and propagates improvements back to the community after iterative refinements.
• Translation Style Customization: A manager uses the Google Docs add-on to apply, augment, and validate translation prompts for culture-specific communications, then shares the enhanced prompt template to support peers seeking context-aware translations.

These scenarios illustrate the productiveness of communal prompt curation, its impact on writing and translation tasks, and the accessibility enabled by tight document-LLM integration (Wang et al., 2024).

6. Technical Formalisms and Algorithms

Wordflow’s core logic can be precisely described as follows:

Prompt Template Expansion:

Let $T$ be a template containing zero or one “{{text}}” placeholders. For input $x$,

$$P = \begin{cases} \mathrm{replace}(T, \{\{text\}\}, x), & \text{if “\{\{text\}\}” in } T \ T \Vert x, & \text{otherwise} \end{cases}$$

where “$\Vert$” denotes string concatenation.

Diff Highlighting:

Given input $x$ and LLM output $y$, Wordflow computes the minimal diff $\Delta$ using Myers’s algorithm, $O(ND)$ complexity, with $N = |x| + |y|$ and $T$0 the edit distance.

Output Postprocessing:

A user-supplied regex $T$1 and replacement $T$2 can extract display segments: $T$3

Prompt Versioning:

While not formalized in the interface, prompt lineage can be viewed abstractly as a DAG $T$4, where $T$5 are prompt versions and $T$6 indicates adaptation from $T$7 to $T$8 (Wang et al., 2024).

7. Implementation, Open Source, and Planned Evaluation

The Wordflow stack includes:

• Frontend: Web Components + Lit Element (PWA)
• LLM inference: MLC-LLM via WebGPU; remote APIs for commercial models
• Backend: AWS API Gateway + DynamoDB
• Storage: IndexedDB (local persistence)
• Google Docs add-on: Embedded Editor View side-panel

Source and deployment:

• Source code: https://github.com/poloclub/wordflow
• Public demo: https://poloclub.github.io/wordflow

Wordflow has been deployed as a design probe; a full-scale user study is planned, structured to log prompt-copying rates, pre-reshare modifications, and to cluster use cases. The study aims to quantify social prompt engineering’s effect on usability and prompt quality (Wang et al., 2024).

---

Wordflow represents a move toward inclusive, community-driven LLM interfaces, enabling prompt engineering through social collaboration and local computation. Its modular architecture, prompt lifecycle, and user-facing design lower technical barriers while promoting best-practice replication and collective refinement (Wang et al., 2024).