Use Thread-of-Thought prompting

​

What is Thread-of-Thought (ThoT) prompting

Thread-of-Thought is an approach that extends chain-of-thought prompting by maintaining a continuous, evolving reasoning process across multiple, related prompts. It’s like having a conversation where each new idea builds on previous ones, helping the LLM to think more deeply and keep track of all the details as we explore a topic.

​

How to implement Thread-of-Thought prompting

1. Provide the original query and context.
2. Use a clear structure. Clearly mark sections with headings, bullet points, or other delimiters for easier parsing.
3. Follow conventions. For instance, use Markdown headings or JSON keys.
4. Use follow-up prompts. Build upon previous thoughts and insights to create a more natural, ongoing thought process.

​

Example

Initial Prompt:

Let’s develop an AI-powered travel planning application. Begin by identifying a key pain point in current travel planning experiences.

Model responds with a challenge in traveling planning, e.g., overwhelming information.

Follow-up:

Great observation! Outline a preliminary concept for an AI travel companion that can address these personalization challenges.

Model suggests what the travel planning platform can offer

Next in thread:

Let’s explore the technological capabilities we need to create such a personalized travel experience. What specific AI and data technologies would power this platform?

Model suggests the technologies needed to create the platform

Continuing:

Consider the user experience and data collection. How would the AI gather and utilize user preferences while maintaining privacy and providing increasing personalization?

Model suggests how the AI can gather and utilize user preferences while maintaining privacy and providing increasing personalization

… The thread continues, building upon previous responses

​

Why use Thread-of-Thought prompting

• Promotes coherent reasoning and logical flow over time.
• Improved context handling builds upon previously established knowledge.
• Better problem decomposition breaks large challenges into manageable steps.
• The model is flexible and adapts its reasoning based on evolving information.
• This approach mimics the natural human-like thought progression.

​

Tips for effective Thread-of-Thought prompting

• Begin with a well-defined initial prompt.
• Encourage referencing of earlier points as needed.
• Periodically summarize key points to maintain focus.
• Regularly filter or refocus context to avoid overload.
• Structured Progression: Move through logical phases of reasoning.
• Allow revision of earlier ideas to refine them.