The Working Set Prompt: Keeping LLM Outputs Consistent Across Multi-Step Work
The article discusses the 'Working Set Prompt' technique to maintain consistent and focused outputs from large language models (LLMs) across multi-step workflows.
Why it matters
The Working Set Prompt technique can help improve the reliability and efficiency of LLM-powered workflows in various domains, from product development to research and analysis.
Key Points
- 1Define a small set of relevant facts (5-10 items) as the 'working set'
- 2Include only the working set in each prompt as a concise reference
- 3Update the working set explicitly when facts change
- 4Helps reduce output drift, improve reproducibility, and optimize token usage
Details
The article introduces the 'Working Set Prompt' technique to address the challenge of maintaining consistent and focused outputs from large language models (LLMs) across multi-step workflows. The key idea is to define a small set of 5-10 relevant facts or context items as the 'working set', and include only this concise reference in each prompt to the LLM. This helps reduce output drift, makes the results more reproducible, and optimizes token usage compared to using a large context window. The author provides an example working set for a product feature development workflow, including elements like the product goal, user story, constraints, current state, and next steps. By explicitly updating the working set as the context changes, the LLM can generate outputs that remain stable and aligned with the evolving task at hand.
No comments yet
Be the first to comment