What are your thoughts on building and optimizing prompts?
I use an informal Google‑sheet approach: I take a prompt, run it on about 10 different inputs, score the outputs, and pick the prompt with the highest score. It’s not perfect, but it’s directionally correct and works well for quick iteration. For getting a voice agent to pause, you can try adding pause tokens like little stars or special symbols; whether it works depends on the specific voice model, as some are trained on audiobooks or transcripts that include pause tokens. It’s not guaranteed every time, but it can help depending on the model you’re using.
Yes, using a voice agent to delay and capture hot leads is a smart, nuanced tactic. For instance, after a form submission, have the agent call within five seconds, acknowledge the form, and say a teammate will connect shortly—this buys time for a human agent while conveying impressive speed and forward‑thinking tech. The agent can simply greet the prospect and promise to patch them through in 10‑15 seconds, or handle basic tasks like scheduling a callback. The key is to be transparent that it’s an AI, not to deceive the prospect into thinking they’re talking to a human.
Don’t use an AI voice agent for speed‑to‑lead calling. People feel disrespected or grossed out when they get an AI call right after submitting a form; the reaction is annoyance, not appreciation. Since form fills are directly convertible to money, annoying them lowers conversion probability. If your funnel represents every dollar you’ll make, cutting conversion by 20 % cuts revenue by 20 %. Voice agents perform worse than humans; you save on staffing but lose 20 % revenue. For example, staffing a setter cost $4 k/month and yielded $20 k, while switching to an AI voice agent dropped cost to $100/month but revenue fell to $5 k. The trade‑off is not worth it.
In most industries, AI voice agents for speed‑to‑lead don’t work well because even a small chance of the prospect realizing it’s an AI destroys trust and willingness to pay. That tiny risk can cost you a significant portion of revenue—often closer to 20‑40%—rather than just saving 1%. The only viable use is a highly nuanced approach: have the prospect fill a form, then immediately show a countdown (We're calling you in 3-2-1) and connect them to a live person, using the AI voice only to announce the imminent call and buy your team a few seconds to prepare. In that scenario the AI acts as a bridge, not the main interlocutor, and can be effective. Outside of that specific setup, AI voice agents are not ready for sales closing or broad lead conversion, and relying on them will likely hurt performance.
One of the big things we're using now is a humanization prompt. I created a statistical model of my tone of voice across all my writing on the internet, then use that model to iteratively produce sentence fragments for content. For each fragment I score the perplexity (statistical deviation) from my own writing and select the lowest‑perplexity options, resulting in content that sounds exactly like me. This lets me generate AI content in my own voice, avoiding the typical AI‑generated feel. You can also use AI to generate drafts and then reformat them to match your tone, which works fairly well despite occasional artifacts that appear only about once every few pages. This dramatically cuts down the time I spend on Q&A and content creation.
Beyond that, I'm thinking about more advanced loops to improve existing automations. For example, you could run a Claude code loop that scans your N8N or Claude routines and tries to make them 10 % more time‑ or cost‑efficient, logging results to a shared file. Over time the system can auto‑research, generate hypotheses, test them, and if it reaches the 10 % improvement target it stops. Running this continuously can re‑jig automation components, reduce module count, convert them into code blocks, and eventually offload them to serverless providers like Modal or Trigger, cutting costs by up to 90 % and making the automations run orders of magnitude faster.