Hey Mark, thanks for reaching out. Totally hear you and prompt engineering to get a desired result across thousands of rows can take some time to hone in on.
First, I noticed your OpenAI account is not on a Tier 2 plan meaning the rate limits will lead to slower loading times and challenges when trying to perform web scraping (more info below)
I've added some extra credits to adjust the prompt a bit. When it comes to this prompt specifically, there are a few areas that we can improve on. The initial question about TAM is relatively subject for an AI to interpret. What may be more effective is asking AI to look at any customers or testimonials featured on their website that can be used as a guide to determine the size of companies they're reaching out to.
As for the order size. Let's say you weren't using AI here, I'm curious how would you find what the average order size is? Is this publicly available on their website or anywhere else that we could look?
Industry fit could be a separate enrichment that may make this process more effective. As for Growth Stage and needs, how would you determine if a company is struggling with lead generation? and have tried cold outreach? or what information on their website would point to this?
Same thing goes for marketing team insights. Employee count is something we should remove and have as a separate enrichment.
The theme here is that a lot of these questions don't have direct data to point to that is publicly and readily accessible for the AI to use to make a decision here. Addittionally, while there are many questions prompted to the AI, there isn't a clear path for the AI to make a decision.
I.e. "Does the company seem to be struggling with predictable growth or lead generation?" even if the answer was easily available, there's not a clear connection between the answer to this question and whether it meets your ICP fit or not.
Overall, the prompt is asking a ton of subjective questions that likely don't have data to point to when making a decision. And without a clear connection between the answers to these questions and how that impacts whether this is an ICP fit or not, the AI is going to struggle.
I would recommend separating out a few of these questions (number of employees, marketing headcount, industry fit) into their own separate enrichments:
* Headcount is an enrichment waterfall we provide
* Headcount of specific departments is another enrichment we provide using our "Find contacts at company by criteria enrichment"
* Industry fit, you can pull the industry this company is by using our "Enrich company from profile" and we can filter out companies that don't match your ICP
To conclude, it's not that these data points aren't able to be found and used to make an ICP determination. However, trying to do them all at once, in a single prompt, without clear instructions of what to do if those answers are found will not leave AI enough context to make reliable decision.
Curious to hear your thoughts here. This task is challenging, and trying to do this at scale means having very specific criteria for success whether you're prompting or using other enrichments.
About Rate Limits
----------------------------
Today, Clay requires 30,000 TPM (tokens per minute) for "ChatGPT - Generate Text" columns and 450,000 TPM for Claygent.
"ChatGPT - Generate Text" will work with any tier, as long as the API key has access/can be used. To use Claygent, you will need a Tier 2 API key from OpenAI, as it supports the required TPM rates.
You can check your OpenAI API usage or limits following these links:
• Usage
• Limits
• Usage Tiers
