Directed Intelligence
Directed Intelligence makes LLMs smarter and more reliable. It combines language models (LLMs) with scripting layers to solve problems we encountered with ChatGPT-style applications.
Why we built it: We started by adding a ChatGPT-style features into our app. It didn't work well:- Users didn't know how to get good results
- Responses were inconsistent and often unreliable
- Long conversations drifted off-topic
- Advice was too generic to be useful
- Experts without technical skills couldn't add their knowledge
- Long conversations still lost focus
- Complex prompts led to tool misuse
- The LLM got confused about what to do next
- Many focused scripts, not one giant prompt
- Experts write scripts in natural language
- LLM selects scripts based on conversation needs
- Structured interactions with room for flexibility
- Active topic management to keep conversations on track
So we built an implementation of this pattern.
How it works:
- Like ChatGPT, it starts with a base language model (LLM) like GPT-4o
- A system selects an appropriate script based on user input
- The script contains step-by-step instructions for the LLM to follow
- LLM follows script, adapting language model outputs
- The system monitors context, switching scripts if needed
- External tools added to LLM input only when script calls for them
- The system maintains conversation history to inform script selection and execution
- Uses many scripts for different product development tasks
- Covers everything from idea validation to sprint planning
- LLM switches between scripts as the conversation evolves
- Maintains overall product development structure
- Gives specific guidance for each task, not generic advice
Directed Intelligence isn't just a band-aid fix. It's a fundamental shift in how we design LLM interactions. If you’re curious to see it in action try the demo at https://fforward.ai.
Example ScriptCameron Westland, Max Cameron, Jon Friis
June 24, 2024