Language Model Prompt Automation for Website Tasks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional automation systems require technical expertise and are not accessible to non-technical users, limiting their usability, while existing natural language processing techniques face challenges in efficiently performing complex tasks on websites due to limitations in prompt size and performance when broken down into multiple steps.
Innovation Solution
A computer-implemented method using language model prompts and templates, where task and subtask prompts are generated from web browser data to guide a language model in performing website tasks, allowing for the breakdown of complex tasks into manageable subtasks and improving performance through few-shot learning techniques.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If traditional automation systems are used, then task automation can be achieved, but technical expertise is required and usability is limited
Solution Approach 1:
The patent replaces traditional programming-based automation mechanisms with a language model-based system. Instead of requiring users to write code or configure complex automation rules, users can simply provide natural language descriptions of desired tasks, and the language model translates these into executable automation actions. This substitution of mechanical programming with linguistic interaction directly resolves the contradiction by eliminating technical expertise requirements while maintaining automation functionality.
Solution Approach 2:
The language model serves as an intermediary between the user's natural language intentions and the automation system's execution capabilities. The system includes a language model that receives user task descriptions in natural language, processes them through prompting techniques, and generates appropriate automation actions. This intermediary layer translates complex automation requirements into simple user-friendly language inputs, thereby improving ease of operation without requiring technical knowledge.
2Ease of operation
If natural language processing techniques are used, then accessibility to non-technical users is improved, but performance on complex tasks is limited due to prompt size constraints
Solution Approach 1:
The patent segments complex web tasks into multiple subtasks, each handled by separate prompt templates. Instead of attempting to handle entire complex tasks within a single prompt, the system breaks down the task hierarchy into atomic actions (e.g., navigation, data entry, form filling), each with its own specialized prompt template. This segmentation allows the language model to process and execute complex tasks reliably by composing smaller, manageable actions together, thereby maintaining both accessibility and task performance.
Solution Approach 2:
The system employs preliminary action through pre-defined prompt templates that are prepared in advance for common task patterns. These templates include example inputs and expected outputs that guide the language model's behavior. By having pre-structured templates ready before execution, the system ensures reliable performance on complex tasks while maintaining user accessibility, as the templates handle the complexity automatically without requiring users to understand or specify detailed parameters.
3Reliability
If prompt templates with few-shot learning are used, then task performance is improved, but system complexity increases
Solution Approach 1:
The patent uses copying through prompt templates that replicate successful task patterns from training data. Each template contains example inputs and outputs that copy proven effective interactions between users and systems. When executing tasks, the language model selects appropriate templates and adapts them to current contexts, effectively copying successful patterns without requiring complex real-time reasoning. This copying mechanism improves task performance by leveraging pre-learned patterns while keeping system complexity manageable, as templates are static and pre-processed rather than dynamically generated.
Data Source
AI summary
A task, such as task completed using a website, may be automated by submitting prompts to a language model and requesting that that language model provide one or more next actions to be performed to complete the task. The accuracy of the language model in providing correct actions may be improved by using one or few-shot learning where examples of completing a task are provided in a prompt to the language model. The accuracy of the language model may also be improved by breaking a task into subtasks. A prompt may be submitted to the language model to request that the language model indicate a subtask to be performed to complete the task. A prompt may then be submitted to the language model to request that the language model indicate a next action to be performed to complete the subtask.