ToT Three-Agent Consensus Title Generator

ToT Loop Article Titles 

This scenario requires you to have access to multiple LLMs. It utilizes Make.com’s native GPT integration, but also makes an API call to OpenRouter (CLICK HERE). OpenRouter is a service that gives you access to most LLMs available today. This scenario uses Claude, but you can easily switch that out for many others. 

This scenario (in its current form) also requires a (free) Jotform account (CLICK HERE). 

Once you have a Jotform account, you can create a form (it’s really easy). Just choose the “short text” option and place it in a new form - 

This form is for taking in a topic. 

After you have your form created in Jotform, you can connect it to the first module. You’ll need to “Create a Webhook” and then choose your new form from the dropdown…

Then, when you fill out your form with a topic, it will trigger the scenario and feed your topic to the A.I. 

This scenario is a simplified version of a prompting program called “Tree of Thought.” 

In the original program, an LLM was asked to give an array of answers to questions, and then another instance of the LLM was asked to crawl the answers and choose the best one. 

This scenario mimics that behavior, but it uses TWO LLMs. 

The scenario is prompted to have GPT offer five article title suggestions. Then, another instance of GPT chooses the best title. 

After that, Claude 2 also chooses the best title. 

You’ll need to input your OpenRouter API key to call Claude -

The scenario loops and repeats up to 10 times OR until both LLMs have mutually agreed on the best title (a new set is regenerated each loop so they always get a fresh batch to choose from). 

With this setup, however, you can prompt it to look for consensus for any body of text (not just titles). 

The magic is in the looping mechanism. 

The scenario is, by default, set to repeat 10 times, but you can adjust that to however many you like. 

Then, the first Filter looks like this 

This SETS the filter to look for the break repeat variable. If it doesn’t exist, the repeater continues. 

The second filter is the LLM match.

See, the LLMs are tasked to only provide the number representing their choice. That way, the numbers can be compared, and the break repeat variable isn’t set until the numbers match. 

This scenario uses TWO different LLMs for fun, but you can also just use GPT with separate system prompts. 

For more information on how to utilize different LLMs in the OpenRouter API call it is here => CLICK HERE

Feel free to modify the LLM prompts, and insert this flow into any of your other automations to leverage LLM Tree of Thought consensus. 

ToT 3rd Agent Article Titles 

This scenario is a simplified version of the first. 

Rather than look for consensus, it simply has a third “agent” choose between two options. 

All the set up is the same, but with this scenario GPT is asked for 5 titles again, then asked to go back and select which title is the BEST. 

Claude is then prompted to do the exact same. 

Then, GPT, with a different system prompt, is asked to pick the better of the two.

You do not have to use GPT for the last decision. It can be Claude, Gemini, or any Openrouter supported LLM. 

ToT 3 Agent Consensus Article Titles 

What? Enough with the Tree of Thought stuff…right? 

This is the last one I promise. I include this one because, it was actually modeled after a recent paper, and it provides one of the simplest, yet most powerful models. 

This one looks for consensus, much like the very first, but rather than looping repeatedly to find one between two “agents” it uses a third agent as a tie breaker. 

So, there are only two potential outcomes, really. Either a title will be chosen twice, or one will be chosen three times. Either way, you’ll always have a winner. 

REMEMBER 

The point of these flows is not to generate blog titles, per se. It is to understand how to use ToT and consensus to choose the BEST body of text, whatever that text is. 

The actual prompts in these modules are bare-bones. You’ll need to flesh them out to your specific use-case. Try to make the outputs the best they can be, and then have A.I. decide on the better outputs. 

Cheers!

‍