$50k/month selling AI automation services to small businesses

Seems like it should work, nice proposal, and good luck to your business.
 
update:

So this is quickly turning into me getting sucked into a wormhole of assistant-based coding of a massive automation network. It's getting out of hand.

I finally finished my reply bot. Here's how it works.

Instantly sends cold email campaign to a lead telling them we made a quick demo of a custom chatbot for their company, and to let us know if they want to see it
If positive response, AI builds the demo with their name/branding/location on a custom domain (ourdomain.com/lead-company-demo) and sends it back (the email also contains a calendar link if they want to book a call to discuss)

It sounds simple, but it was quite a lot of work building it with AI and getting it to make the demos dynamically. My goal was to be able to feed a massive list of leads in different industries into the campaign and then just let it rip without having to split for dentists, agencies, lawyers, etc. Just mix and match and the bot will send a demo tailored to that industry with nothing hardcoded, all created by AI on the spot. Obviously to get this done took a lot of fine tuning and work, made more difficult by the fact that building with ai agents is like working with an overeager supergenius with a terrible short term memory.

Anyway, once I finished my first project I was all happy and proud of myself. One of the things you have to do when you code like this is tell the bot to make lots of documentation. It doesn't remember from chat to chat, so instead of me trying to explain over and over what I want, I can just tell it once to create a CHANGELOG doc, for example, which will track all recent changes to the code. So then if I open a new chat, I can say "ok go read the changelog so you know where we're at."

In addition to the changelog, while I was building the project I also created more documentation. For example here is a snippet of another doc that is a more high level overview:

# COMPREHENSIVE PROJECT KNOWLEDGE BASE
*Reusable Patterns, Methodologies & Frameworks for Future Projects*

## **MAJOR UPDATE**: ALL-PURPOSE DYNAMIC INDUSTRY SYSTEM
*This project successfully transformed from a hardcoded solar-only system to a dynamic all-industry lead generation platform. This knowledge base now includes patterns for building industry-agnostic AI systems.*

## **TABLE OF CONTENTS**

1. [5 Core Documentation Framework](#5-core-documentation-framework)
2. [Dynamic Industry AI System Architecture](#dynamic-industry-ai-system-architecture)
3. [TaskMaster Integration Methodology](#taskmaster-integration-methodology)
4. [Debugging Methodology & 30-Minute Rule](#debugging-methodology--30-minute-rule)
5. [Cursor Rules Structure & Self-Improvement](#cursor-rules-structure--self-improvement)
6. [Development Workflow Patterns](#development-workflow-patterns)
7. [Project Architecture Patterns](#project-architecture-patterns)
8. [Environment & Configuration Management](#environment--configuration-management)
9. [Testing & Validation Approaches](#testing--validation-approaches)

All of this was written with AI. I dont' remember what I told it, but I probably said something like, "analyze the entire codebase and make some documentation that identifies all the patterns and methodologies we used to build this." What's cool is that the dutiful little AI went and did it. I am not a master programmer, in fact I've never coded anything the old fashioned way. But if you can articulate what you want the AI to do properly, then it can do it for you.

Long story short, after a few days of this kind of thing, I had around 10 documents that all explained different strategies and frameworks for the project. I can share them if there is any interest, I think they are pretty valuable. My whole reason for making them was so that if I wanted to make something else, like make the lead generation system better or have more features, that I wouldn't have to beat the AI with a stick to get it to do what I wanted. I would just be able to feed it some docs and say, "ok go do this now." That's what it's supposed to do in theory, but it's not so easy.

AI Agents

At this point I saw something on yt about people building workflows with AI agents that talk to each other. So I put two and two together and thought, "how can I use these design principles to build a team of agents that will build software for me in a layered approach based on these design principles?"

One thing lead to another, and now I have 3/7 "meta agents" who will each be responsible for 1 step in the process of building future tools. Is it overkill? Probably. Do I need to do it to make stuff? No. But it's very fun.

So bottom line here is the new plan:

Finish building the meta agents (done by Fri)
Use the meta agents to build additional agents for the leadgen workflow (done by end of next week)
Use those agents + my cold email + autodemo to crank out targeted campaigns, generate leads, and close deals (Aug 1 let's say)
THEN start hiring sales people and training them (Sept 1, after 1 month of me working the leads myself)

Happy to share whatever details, I know it seems like it's taken a turn but CC is very fun and I highly recommend you guys get into it.
 
Are you going to build a website for your chat bot/AI services, I would personally want to have a look at a website see some LinkedIn profile ect with some social proof on ect before buying a high ticket service
 
Last edited:
update:

So this is quickly turning into me getting sucked into a wormhole of assistant-based coding of a massive automation network. It's getting out of hand.

I finally finished my reply bot. Here's how it works.

Instantly sends cold email campaign to a lead telling them we made a quick demo of a custom chatbot for their company, and to let us know if they want to see it
If positive response, AI builds the demo with their name/branding/location on a custom domain (ourdomain.com/lead-company-demo) and sends it back (the email also contains a calendar link if they want to book a call to discuss)

It sounds simple, but it was quite a lot of work building it with AI and getting it to make the demos dynamically. My goal was to be able to feed a massive list of leads in different industries into the campaign and then just let it rip without having to split for dentists, agencies, lawyers, etc. Just mix and match and the bot will send a demo tailored to that industry with nothing hardcoded, all created by AI on the spot. Obviously to get this done took a lot of fine tuning and work, made more difficult by the fact that building with ai agents is like working with an overeager supergenius with a terrible short term memory.

Anyway, once I finished my first project I was all happy and proud of myself. One of the things you have to do when you code like this is tell the bot to make lots of documentation. It doesn't remember from chat to chat, so instead of me trying to explain over and over what I want, I can just tell it once to create a CHANGELOG doc, for example, which will track all recent changes to the code. So then if I open a new chat, I can say "ok go read the changelog so you know where we're at."

In addition to the changelog, while I was building the project I also created more documentation. For example here is a snippet of another doc that is a more high level overview:



All of this was written with AI. I dont' remember what I told it, but I probably said something like, "analyze the entire codebase and make some documentation that identifies all the patterns and methodologies we used to build this." What's cool is that the dutiful little AI went and did it. I am not a master programmer, in fact I've never coded anything the old fashioned way. But if you can articulate what you want the AI to do properly, then it can do it for you.

Long story short, after a few days of this kind of thing, I had around 10 documents that all explained different strategies and frameworks for the project. I can share them if there is any interest, I think they are pretty valuable. My whole reason for making them was so that if I wanted to make something else, like make the lead generation system better or have more features, that I wouldn't have to beat the AI with a stick to get it to do what I wanted. I would just be able to feed it some docs and say, "ok go do this now." That's what it's supposed to do in theory, but it's not so easy.

AI Agents

At this point I saw something on yt about people building workflows with AI agents that talk to each other. So I put two and two together and thought, "how can I use these design principles to build a team of agents that will build software for me in a layered approach based on these design principles?"

One thing lead to another, and now I have 3/7 "meta agents" who will each be responsible for 1 step in the process of building future tools. Is it overkill? Probably. Do I need to do it to make stuff? No. But it's very fun.

So bottom line here is the new plan:

Finish building the meta agents (done by Fri)
Use the meta agents to build additional agents for the leadgen workflow (done by end of next week)
Use those agents + my cold email + autodemo to crank out targeted campaigns, generate leads, and close deals (Aug 1 let's say)
THEN start hiring sales people and training them (Sept 1, after 1 month of me working the leads myself)

Happy to share whatever details, I know it seems like it's taken a turn but CC is very fun and I highly recommend you guys get into it.
Amazing update, just a few questions:

- Any chance you could share with us the workflows you have been developing for autoresponse and creation of the custom bot for the customers?

- what are the specific tasks you’re building each of the meta-agents for?

- in the end will you be selling a custom chatbot to the leads? Or still thinking about the idea about the database reactivation?

- will you design the client product yourself? Or will you hire freelancers for it?

Many thanks! Keep up the good work!
 
Are you going to build a website for your chat bot/AI services, I would personally want to have a look at a website see some LinkedIn profile ect with some social proof on ect before buying a high ticket service
I have a landing page set up now but it's very sparse on the details, it's basically a giant reskinned Calendly page. In terms of linkedin, "results" and testimonials and all that, yeah that's not a bad idea. I could have AI make me some fake testimonials or wahtever, but I think what I am going to do instead once I turn my focus more towards sales is to create organic social media content showcasing the workflows.

1753318619432.png

I've seen a lot of people find success on social media by sharing tutorials, tips, tricks, workflows, etc. It's actually a great niche because there is so much interest, and because it's a new field, there's a ton to talk about and all of the info is recent. New breakthroughs being released on the daily basically.

But what I want to do first is finish building this AI project for the meta agents and future lead gen agents. Once I've done that, turned it on, and verified that it works, I'll start cranking social.
Amazing update, just a few questions:

- Any chance you could share with us the workflows you have been developing for autoresponse and creation of the custom bot for the customers?

- what are the specific tasks you’re building each of the meta-agents for?

- in the end will you be selling a custom chatbot to the leads? Or still thinking about the idea about the database reactivation?

- will you design the client product yourself? Or will you hire freelancers for it?

Many thanks! Keep up the good work!
Yes, I could share the workflows. I thought about sharing the github repo as it's all in there, but I think I may develop this into a commercial solution down the line so idk about having all the code out there. Maybe if you ask me really nicely :)

Specific tasks for the meta agents are as follows:
  • Input Funnel → Clean the idea, check what exists, ask missing questions, output a clear brief.
-this is technically not an agent, it's just hte input that would be given to the system to start it off. So think of it like a prompt or whatever. "Build me a lead generation system for a dentist in California." For this, I need to make it complex like an intake form so that the team of agents doesn't just assume what they want. So we'd have to have a list of 20-30 short questions in the input funnel, i.e.

"what do you want this to do?"
"who is this for?"
"what do you want this to not do?"
"do you want to use any special software"

I haven't done this yet, but once the inputs are created, here is how they are handled.

Project Planner (PRD Parser & Task Graph Builder) - most important

Reads the brief and turns it into a to‑do list with clear steps, dependencies, and pass/fail checks. Updates it when the brief changes.

Code Factory (Template Engine / Code Factory Agent)
Writes the actual code, tests, and config files from reusable templates. If tasks change, it regenerates so things stay in sync.

Health Checker (Infrastructure Orchestrator Agent)
Scans the repo for ticking time bombs like hardcoded limits or missing env vars. Flags them and suggests safer fixes before deployment.

Data Wiring Tech (Parameter Flow Agent)
Lists every input and output between services, makes sure names and types match, and auto‑builds tests to prove the connections work.

Living Docs Keeper (Five Document Framework Agent)
Maintains five plain‑English docs: what we are building, why it matters, how it is built, how to run or fix it, and what changed. Updates them automatically.

Stall Timer (Thirty Minute Rule Agent)
Watches for tasks that are stuck. When the timer hits the limit, it triggers a simpler approach, skips the task for now, or pings a human. Logs what happened.

Project Starter Kit (Scaffold Generator / All Purpose Pattern Agent)

Spins up a fresh repo with your standard folder layout, CI, linting, tests, and infra stubs so new projects start clean and consistent.

Deployer (Vercel Native Architecture Agent)
Sets up the Vercel project, environment variables, routes, logging, and monitoring, then deploys in the same reliable way every time.

Research Feed (R and D Agent)
Constantly hunts for better libraries, examples, and patterns, and feeds that back into templates, prompts, and task plans so the whole system improves over time.

Why is the Parser most important? Because it breaks your project into manageable tasks and subtasks, then uses Perplexity to accurately research the most accurate solution for accomplishing that task. I use this method to build everything, and if it's not working I refuse to proceed. It's 1 million times better than just relying on the AI to whip up some code off the top of its head. Extra planning = smooth project.

Many of these agents are not things that you would consider when building anything, but they were constructed because the processes they automate were repeated enough times in this project that it's just impossible to deny how important they are. There are so many components that go into writing software that you will end up creating these little tools and automations inside the greater automation just because it simplifies an important part of the process.

For example, something that's not on this list is a Parameter Matching document. The currently spans multiple "surfaces" each with their own sets of variables, arguments, documentation, etc. It's a lot of shit to keep track of. If I am working on Instantly, N8N, my own custom API, Vercel, and a vector database, chances are different variables are mapped differently.

This can be a big problem when you have thousands of lines of code and variables across 8 surfaces that need to be tracked so they can be referenced later in the project. One little error means the whole thing breaks.

The Parameter Matching document is like a giant glossary of terms for how variables and arguments are coded on the different surfaces. So any time the AI is writing code, it can reference that doc and know how it's supposed to refer to specific items. And since parameter matching is important enough that we created documentation on it, that means it's important enough to turn into an agent. Here is a snippet of the parameter flow agent:

1753318761390.png

Little things like that will pop up after going in circles on the same error for 2 hours. I once spent an entire day on a bug because I was calling the wrong arguments for the Instantly API. And what makes matters worse is that the AI will forget what's worked in the past and revert to making the same mistakes it made. So you have to constantly remind it to do the same thing.

In fact, the short term memory of AI coding assistants is a massive roadblock for getting momentum while working on a project. There is no inherent "memory" in the sense that you can store files that will persist after you restart the coding agent. Had a bunch of good breakthroughs in your last chat? Well they're gone once you close the chat.

The solution I've found is to create something called a RAG, which is a database where you can store tagged information about shit that's important in your project. So if you say, "ok I want you to continue building the parser agent with taskmaster and context7," then it will use the RAG to look up what those terms mean, how to use them, and there will be very little guesswork from the AI as to what to do. Combine this with the actual parser agent that uses Perplexity to research each little task and subtask for whatever you're working on, and you have a shot at being efficient.

Anyway, the project is moving along. I am actually wrapping up all the meta agents now. But I had to stop to create a RAG. And I also added an "observational monitoring" dashboard as well.

This is still in the process of being worked on, but basically it's a big dashboard that will show you in real time what all of your agents are doing. If you've read this far then you understand that this project is complex, lots of moving parts, lots of places for the chain to break. And if the chain breaks, we don't always know where it broke. But with this observational dashboard, all of the agents are monitored and we can see exactly what they're doing. So it's going to be much easier to fine tune the development.

1753318576282.png

I honestly think I have the seed of a commercial software development workflow that will be able to spit out very solid tech in a more or less automated fashion. All of this was done to make it easier to build tools for my lead generation suite, but I think long term this meta agent suite is actually going to be what's most valuable.

My previous lead campaign is out of juice so I will need to get more leads and crank through again. I should probably just go for the direct appraoch this time and not try to book podcasts. We will see
 
I have a landing page set up now but it's very sparse on the details, it's basically a giant reskinned Calendly page. In terms of linkedin, "results" and testimonials and all that, yeah that's not a bad idea. I could have AI make me some fake testimonials or wahtever, but I think what I am going to do instead once I turn my focus more towards sales is to create organic social media content showcasing the workflows.

View attachment 461148

I've seen a lot of people find success on social media by sharing tutorials, tips, tricks, workflows, etc. It's actually a great niche because there is so much interest, and because it's a new field, there's a ton to talk about and all of the info is recent. New breakthroughs being released on the daily basically.

But what I want to do first is finish building this AI project for the meta agents and future lead gen agents. Once I've done that, turned it on, and verified that it works, I'll start cranking social.

Yes, I could share the workflows. I thought about sharing the github repo as it's all in there, but I think I may develop this into a commercial solution down the line so idk about having all the code out there. Maybe if you ask me really nicely :)

Specific tasks for the meta agents are as follows:
  • Input Funnel → Clean the idea, check what exists, ask missing questions, output a clear brief.
-this is technically not an agent, it's just hte input that would be given to the system to start it off. So think of it like a prompt or whatever. "Build me a lead generation system for a dentist in California." For this, I need to make it complex like an intake form so that the team of agents doesn't just assume what they want. So we'd have to have a list of 20-30 short questions in the input funnel, i.e.

"what do you want this to do?"
"who is this for?"
"what do you want this to not do?"
"do you want to use any special software"

I haven't done this yet, but once the inputs are created, here is how they are handled.

Project Planner (PRD Parser & Task Graph Builder) - most important
Reads the brief and turns it into a to‑do list with clear steps, dependencies, and pass/fail checks. Updates it when the brief changes.

Code Factory (Template Engine / Code Factory Agent)
Writes the actual code, tests, and config files from reusable templates. If tasks change, it regenerates so things stay in sync.

Health Checker (Infrastructure Orchestrator Agent)
Scans the repo for ticking time bombs like hardcoded limits or missing env vars. Flags them and suggests safer fixes before deployment.

Data Wiring Tech (Parameter Flow Agent)
Lists every input and output between services, makes sure names and types match, and auto‑builds tests to prove the connections work.

Living Docs Keeper (Five Document Framework Agent)
Maintains five plain‑English docs: what we are building, why it matters, how it is built, how to run or fix it, and what changed. Updates them automatically.

Stall Timer (Thirty Minute Rule Agent)
Watches for tasks that are stuck. When the timer hits the limit, it triggers a simpler approach, skips the task for now, or pings a human. Logs what happened.

Project Starter Kit (Scaffold Generator / All Purpose Pattern Agent)
Spins up a fresh repo with your standard folder layout, CI, linting, tests, and infra stubs so new projects start clean and consistent.

Deployer (Vercel Native Architecture Agent)
Sets up the Vercel project, environment variables, routes, logging, and monitoring, then deploys in the same reliable way every time.

Research Feed (R and D Agent)
Constantly hunts for better libraries, examples, and patterns, and feeds that back into templates, prompts, and task plans so the whole system improves over time.

Why is the Parser most important? Because it breaks your project into manageable tasks and subtasks, then uses Perplexity to accurately research the most accurate solution for accomplishing that task. I use this method to build everything, and if it's not working I refuse to proceed. It's 1 million times better than just relying on the AI to whip up some code off the top of its head. Extra planning = smooth project.

Many of these agents are not things that you would consider when building anything, but they were constructed because the processes they automate were repeated enough times in this project that it's just impossible to deny how important they are. There are so many components that go into writing software that you will end up creating these little tools and automations inside the greater automation just because it simplifies an important part of the process.

For example, something that's not on this list is a Parameter Matching document. The currently spans multiple "surfaces" each with their own sets of variables, arguments, documentation, etc. It's a lot of shit to keep track of. If I am working on Instantly, N8N, my own custom API, Vercel, and a vector database, chances are different variables are mapped differently.

This can be a big problem when you have thousands of lines of code and variables across 8 surfaces that need to be tracked so they can be referenced later in the project. One little error means the whole thing breaks.

The Parameter Matching document is like a giant glossary of terms for how variables and arguments are coded on the different surfaces. So any time the AI is writing code, it can reference that doc and know how it's supposed to refer to specific items. And since parameter matching is important enough that we created documentation on it, that means it's important enough to turn into an agent. Here is a snippet of the parameter flow agent:

View attachment 461150

Little things like that will pop up after going in circles on the same error for 2 hours. I once spent an entire day on a bug because I was calling the wrong arguments for the Instantly API. And what makes matters worse is that the AI will forget what's worked in the past and revert to making the same mistakes it made. So you have to constantly remind it to do the same thing.

In fact, the short term memory of AI coding assistants is a massive roadblock for getting momentum while working on a project. There is no inherent "memory" in the sense that you can store files that will persist after you restart the coding agent. Had a bunch of good breakthroughs in your last chat? Well they're gone once you close the chat.

The solution I've found is to create something called a RAG, which is a database where you can store tagged information about shit that's important in your project. So if you say, "ok I want you to continue building the parser agent with taskmaster and context7," then it will use the RAG to look up what those terms mean, how to use them, and there will be very little guesswork from the AI as to what to do. Combine this with the actual parser agent that uses Perplexity to research each little task and subtask for whatever you're working on, and you have a shot at being efficient.

Anyway, the project is moving along. I am actually wrapping up all the meta agents now. But I had to stop to create a RAG. And I also added an "observational monitoring" dashboard as well.

This is still in the process of being worked on, but basically it's a big dashboard that will show you in real time what all of your agents are doing. If you've read this far then you understand that this project is complex, lots of moving parts, lots of places for the chain to break. And if the chain breaks, we don't always know where it broke. But with this observational dashboard, all of the agents are monitored and we can see exactly what they're doing. So it's going to be much easier to fine tune the development.

View attachment 461147

I honestly think I have the seed of a commercial software development workflow that will be able to spit out very solid tech in a more or less automated fashion. All of this was done to make it easier to build tools for my lead generation suite, but I think long term this meta agent suite is actually going to be what's most valuable.

My previous lead campaign is out of juice so I will need to get more leads and crank through again. I should probably just go for the direct appraoch this time and not try to book podcasts. We will see
Awesome, thank you again for the hard work and posting your updates and answers to the questions! Following up closely and replicating some of the ideas for my AI side hustle too, here are some of the questions that come to mind:

  • Well this is not a question, but I just wanted to ask really nicely if you could share the workflows or the GitHub repo, I could really use it for my side hustle and I believe I could provide you with some constructive feedback too (but please know I’m not a great programmer neither) and could bring some new ideas to the table.
  • 1. Do you already have some leads that replied and scheduled a call? Do you already have a strategy for the pricing (i.e. one time project or offer a subscription)?
  • 2. How are you keeping track of the cost and future invoices? – Might look like an easy one but I tried everything and at the end of the day I always end up using a simple excel sheet, just wondering if there might be a tool I still do not know.
Once again, thank you!
 
Solid direction and honest breakdown respect for documenting the full process. I’d suggest doubling down on one high-converting offer (like database reactivation) while refining your outreach copy. Also, consider hiring a part-time content creator to kickstart your social presence without draining your energy. Keep pushing!
 
so essentially , 1. chatbots to contact dead leads to try to reactivate, 2. chatbot for website 3. chatbot for prospective leads to cold email and convert them into appointments. 1 and 3 are similar , 2. is probably widely used and competitive, but uniqueness will be your selling point. Seems that persuading the small business to pay you is a big hurdle.
 
solid pivot, and the breakdown shows ur thinkin straight. ai automation for smb is hot rn, and u got the cashflow buffer to experiment hard without panic. goodluck with the journey
 
I have a landing page set up now but it's very sparse on the details, it's basically a giant reskinned Calendly page. In terms of linkedin, "results" and testimonials and all that, yeah that's not a bad idea. I could have AI make me some fake testimonials or wahtever, but I think what I am going to do instead once I turn my focus more towards sales is to create organic social media content showcasing the workflows.

View attachment 461148

I've seen a lot of people find success on social media by sharing tutorials, tips, tricks, workflows, etc. It's actually a great niche because there is so much interest, and because it's a new field, there's a ton to talk about and all of the info is recent. New breakthroughs being released on the daily basically.

But what I want to do first is finish building this AI project for the meta agents and future lead gen agents. Once I've done that, turned it on, and verified that it works, I'll start cranking social.

Yes, I could share the workflows. I thought about sharing the github repo as it's all in there, but I think I may develop this into a commercial solution down the line so idk about having all the code out there. Maybe if you ask me really nicely :)

Specific tasks for the meta agents are as follows:
  • Input Funnel → Clean the idea, check what exists, ask missing questions, output a clear brief.
-this is technically not an agent, it's just hte input that would be given to the system to start it off. So think of it like a prompt or whatever. "Build me a lead generation system for a dentist in California." For this, I need to make it complex like an intake form so that the team of agents doesn't just assume what they want. So we'd have to have a list of 20-30 short questions in the input funnel, i.e.

"what do you want this to do?"
"who is this for?"
"what do you want this to not do?"
"do you want to use any special software"

I haven't done this yet, but once the inputs are created, here is how they are handled.

Project Planner (PRD Parser & Task Graph Builder) - most important
Reads the brief and turns it into a to‑do list with clear steps, dependencies, and pass/fail checks. Updates it when the brief changes.

Code Factory (Template Engine / Code Factory Agent)
Writes the actual code, tests, and config files from reusable templates. If tasks change, it regenerates so things stay in sync.

Health Checker (Infrastructure Orchestrator Agent)
Scans the repo for ticking time bombs like hardcoded limits or missing env vars. Flags them and suggests safer fixes before deployment.

Data Wiring Tech (Parameter Flow Agent)
Lists every input and output between services, makes sure names and types match, and auto‑builds tests to prove the connections work.

Living Docs Keeper (Five Document Framework Agent)
Maintains five plain‑English docs: what we are building, why it matters, how it is built, how to run or fix it, and what changed. Updates them automatically.

Stall Timer (Thirty Minute Rule Agent)
Watches for tasks that are stuck. When the timer hits the limit, it triggers a simpler approach, skips the task for now, or pings a human. Logs what happened.

Project Starter Kit (Scaffold Generator / All Purpose Pattern Agent)
Spins up a fresh repo with your standard folder layout, CI, linting, tests, and infra stubs so new projects start clean and consistent.

Deployer (Vercel Native Architecture Agent)
Sets up the Vercel project, environment variables, routes, logging, and monitoring, then deploys in the same reliable way every time.

Research Feed (R and D Agent)
Constantly hunts for better libraries, examples, and patterns, and feeds that back into templates, prompts, and task plans so the whole system improves over time.

Why is the Parser most important? Because it breaks your project into manageable tasks and subtasks, then uses Perplexity to accurately research the most accurate solution for accomplishing that task. I use this method to build everything, and if it's not working I refuse to proceed. It's 1 million times better than just relying on the AI to whip up some code off the top of its head. Extra planning = smooth project.

Many of these agents are not things that you would consider when building anything, but they were constructed because the processes they automate were repeated enough times in this project that it's just impossible to deny how important they are. There are so many components that go into writing software that you will end up creating these little tools and automations inside the greater automation just because it simplifies an important part of the process.

For example, something that's not on this list is a Parameter Matching document. The currently spans multiple "surfaces" each with their own sets of variables, arguments, documentation, etc. It's a lot of shit to keep track of. If I am working on Instantly, N8N, my own custom API, Vercel, and a vector database, chances are different variables are mapped differently.

This can be a big problem when you have thousands of lines of code and variables across 8 surfaces that need to be tracked so they can be referenced later in the project. One little error means the whole thing breaks.

The Parameter Matching document is like a giant glossary of terms for how variables and arguments are coded on the different surfaces. So any time the AI is writing code, it can reference that doc and know how it's supposed to refer to specific items. And since parameter matching is important enough that we created documentation on it, that means it's important enough to turn into an agent. Here is a snippet of the parameter flow agent:

View attachment 461150

Little things like that will pop up after going in circles on the same error for 2 hours. I once spent an entire day on a bug because I was calling the wrong arguments for the Instantly API. And what makes matters worse is that the AI will forget what's worked in the past and revert to making the same mistakes it made. So you have to constantly remind it to do the same thing.

In fact, the short term memory of AI coding assistants is a massive roadblock for getting momentum while working on a project. There is no inherent "memory" in the sense that you can store files that will persist after you restart the coding agent. Had a bunch of good breakthroughs in your last chat? Well they're gone once you close the chat.

The solution I've found is to create something called a RAG, which is a database where you can store tagged information about shit that's important in your project. So if you say, "ok I want you to continue building the parser agent with taskmaster and context7," then it will use the RAG to look up what those terms mean, how to use them, and there will be very little guesswork from the AI as to what to do. Combine this with the actual parser agent that uses Perplexity to research each little task and subtask for whatever you're working on, and you have a shot at being efficient.

Anyway, the project is moving along. I am actually wrapping up all the meta agents now. But I had to stop to create a RAG. And I also added an "observational monitoring" dashboard as well.

This is still in the process of being worked on, but basically it's a big dashboard that will show you in real time what all of your agents are doing. If you've read this far then you understand that this project is complex, lots of moving parts, lots of places for the chain to break. And if the chain breaks, we don't always know where it broke. But with this observational dashboard, all of the agents are monitored and we can see exactly what they're doing. So it's going to be much easier to fine tune the development.

View attachment 461147

I honestly think I have the seed of a commercial software development workflow that will be able to spit out very solid tech in a more or less automated fashion. All of this was done to make it easier to build tools for my lead generation suite, but I think long term this meta agent suite is actually going to be what's most valuable.

My previous lead campaign is out of juice so I will need to get more leads and crank through again. I should probably just go for the direct appraoch this time and not try to book podcasts. We will see
so are you a software engineer? or just breaking down the coding assignments into smaller tasks for AI to do the coding? I have created a few applications with AI that work but are small/simple. It would be intuitive that AI can handle smaller tasks better than one huge coding project.
 
so are you a software engineer? or just breaking down the coding assignments into smaller tasks for AI to do the coding? I have created a few applications with AI that work but are small/simple. It would be intuitive that AI can handle smaller tasks better than one huge coding project.
No, not a software engineer. I think of myself more as an entrepreneur since I run an actual business, but I want to pivot. And you're right about AI being better at the small things. But what's appealing to me is that as someone who isn't a programmer but who is always thinking strategy, I can tell an AI coding assistant to make me some software and it will do it.

Maybe I blah blahed too much in my last post, but what I am putting together is a framework for software creation that is procedural and layered. It will take a PRD and spin it up into a working piece of software. That is the goal at least, the trick is to refine the automations and integrations between the automations until they are working together. That's also the challenge though, because you have to be able to sit and think about the logic of how it all fits together.

This project I'm working on now is evolving every day because every day I am working on new things. My process is basically like this:

Decide what I want to do
Bot researches the best way to do it with perplexity
Bot makes a to do list (tasks), broken into subtasks
Provide a gateway to up to date comprehensive code libraries in every language
Bot does its thing and makes V1
I ask the bot to test
Bot tests and finds errors
I ask the bot to iterate, debug, test on its own and come back to me when its doing what I want it to do
Bot writes docs that explain the feature we created and how to use it
Go back to step 1 if I think of something else I want it to do (in the same section of the project)
If not, move onto next feature

At that point, I make a decision if I want to "agentify" the feature we just made and turn it into a meta agent. This is the philsophy that lead me to create the project this way.

For example, today I did a massive reorganization of the codebase and files since it was a total mess. Using the process above, here's more or less what we did:

Decide I want to reoganize the project
Bot researches best way to reorganize a codebase + docs
Bot makes a to do list (tasks), broken into subtasks
....
...
...
Docs created explaining the reasoning and strategy behind how the project was organized.

Will I make a meta agent that organizes projects? Maybe, But if I am just going to build for me for now then it's maybe not such an urgent improvement because my shit's already organized and will theoretically stay organized because of the doc.

But it is possible I may want to "agentify" another set of processes down the line, and that's how the project grows.

Tomorrow I am going to use the meta agents (11 now) to build some of the lead generation tools for me. I have no idea how that is going to go. I have done a minor test of the meta agents yesterday and it actually created something that worked. My idea was a tool that searches YouTube every day for videos based on specific keywords (ex: "how to automate your social media posts with n8n"), reads the video transcript, and searches for relevant projects on github. These github projects and YT transcripts would be parsed with AI into a RAG (like a memory bank), so the end result would be that you have lots of code snippets that build cool projects that you could reference with a coding agent.

The bot would run daily, scraping/parsing transcripts and github repos and throwing the useful bits into a persistant memory. You can then connect that memory to a chatbot that will reference it when you ask it questions. So let's say I accelerate this and run it for 500 videos about how to do stuff with n8n, make.com, claude code, cursor, whatever, then it will return 500 tutorial transcripts, parse those into lord knows how many github repos, throw the code in the RAG, and then theoretically you will have a coding AI assistant who is trained on actual working github repos and YouTube tutorials.

It sounds a little convoluted but once you get bit by the AI coding bug, once you really "get it" that you can create enterprise level software that will solve whatever problem you want as long as you can articulate it well enough for the AI to understand, as long as you are willing to learn how to supply all the layers needed to make big projects (databases, security, servers, front end, RAG, docs, commenting, etc), then you can produce insane working software that is arguably better than most well known commercial solutions for most products.

Take Instantly for example. Nothing against them, they're lovely, I'm a customer of theirs. But they don't provide anything that you can't create yourself with an AI coding agent. It's an interface to send emails, it's not some crazy thing. If you know what you're doing you could make it in a week or two and save yourself $100/mo, fix all their bugs and shit that they don't have, add features you want, and then you have a better working solution.

Doing that may not get you excited, but I'm sure you have some problem in your business that you would like to solve. Well if you feed that problem to an AI coding assistant it may lead you to solving it with an actual tool you can use.

Awesome, thank you again for the hard work and posting your updates and answers to the questions! Following up closely and replicating some of the ideas for my AI side hustle too, here are some of the questions that come to mind:

  • Well this is not a question, but I just wanted to ask really nicely if you could share the workflows or the GitHub repo, I could really use it for my side hustle and I believe I could provide you with some constructive feedback too (but please know I’m not a great programmer neither) and could bring some new ideas to the table.
  • 1. Do you already have some leads that replied and scheduled a call? Do you already have a strategy for the pricing (i.e. one time project or offer a subscription)?
  • 2. How are you keeping track of the cost and future invoices? – Might look like an easy one but I tried everything and at the end of the day I always end up using a simple excel sheet, just wondering if there might be a tool I still do not know.
Once again, thank you!
Idk if I want to share them just yet. I think I may have something really valuable in a few weeks/months. And from my cold email campaign I got around 10-15 responses that I could have turned into phone appointments, and one person who actually scheduled on the calendar. But I never pursued it. And that was only sending around 400 emails a day. I can now send 4000/day, but I am busy working on this AI stuff so I haven't really started doing outreach. Tomorrow I will build more lead gen tools, so maybe once I have a better idea for how those work I will get excited about selling again. Selling DB reactivation doesn't get me so excited, but it does sell well from what i hear.
 
Your clear focus on building a strong foundation in marketing, sales, and fulfillment is the perfect approach to scaling a successful AI automation business—excited to see your progress
 
This post is very long. If you are following this thread and genuinely interested in AI software development, then I suggest you read it all. It's pretty insane what I've been able to create in around a week.

If you don't want to read it all, basically I have a full system of specialized agents (factories) that build software automatically off of a PRD. That's number one. Number two now is that the next thing on the horizon is to build out specialized factories for specific niches. Below I've pasted two outputs from my coding assistant, one to summarize what I've done so far, and the 2nd one is an explanation of the Lead Generation Factory that I'm going to build next. Everything below this was generated by my coding assistant.


***

What I Built - Simple Explanation

I created a system where AI robots build other AI robots, then those robots work together to build complete software projects automatically.

---
What This Actually Is

Think of it like this: Instead of hiring 5 different contractors (backend developer, frontend developer, etc.), I built 5 AI robots that do those jobs.
Then I built a factory that creates and manages these robots.

The crazy part: The robots talk to each other and coordinate their work automatically. No human management needed.

---
The Robot Factory (Meta-Agent Factory)

What it does: Creates specialized AI robots for different jobs
- Like having a robot factory that can build mechanic robots, electrician robots, plumber robots, etc.
- 11 different types of builder robots that know how to create workers
- Zero limitations - can adapt to build robots for ANY type of project

Real example: Tell it "I need a robot that builds websites" → Factory creates a website-building robot in minutes

---
️ The 5 Worker Robots I Built

Backend Robot ️

- Builds databases and server systems
- Creates APIs (how apps talk to servers)
- Handles security and user accounts

Frontend Robot

- Builds user interfaces (what people see and click)
- Makes websites look good and work on phones
- Ensures apps are accessible for disabled users

DevOps Robot

- Handles putting apps on the internet
- Sets up automated testing and deployment
- Monitors if everything is running properly

QA Robot

- Creates tests to make sure nothing breaks
- Finds edge cases and potential problems
- Generates comprehensive testing plans

Documentation Robot


- Writes instruction manuals for the code
- Creates API documentation
- Builds help guides and tutorials

---
The Coordination System (UEP)

The magic part: All robots communicate through a "universal translator"
- Backend robot tells Frontend robot: "I built these APIs for you to use"
- DevOps robot tells everyone: "I set up the servers, everything's ready"
- QA robot tells everyone: "I found these issues that need fixing"

Result: They work together like a perfect team, no confusion or miscommunication

---
The Memory System (RAG)

What it does: Remembers everything from every project
- 659+ files of knowledge stored and searchable
- Never forgets solutions to problems
- Each new project builds on everything learned before

Think of it like: A company where every employee has perfect memory and shares all knowledge instantly

---
Mission Control Dashboard

Visual monitoring system where you can watch all the robots working
- See which robots are busy, which are idle
- Monitor progress on complex projects
- Real-time status updates on everything happening

It's like: Air traffic control, but for AI robots building software

---
What This Actually Accomplishes

Before: Want to build an app? Hire backend dev ($100k/year), frontend dev ($90k/year), DevOps engineer ($120k/year), QA tester ($70k/year), technical
writer ($60k/year) = $440k/year team

After: Tell my robot factory: "Build me an app with user accounts, payment processing, and mobile interface"

The robots automatically:
1. Backend robot creates the database and payment APIs
2. Frontend robot builds the mobile-friendly interface
3. DevOps robot puts it online with automatic backups
4. QA robot tests everything thoroughly
5. Documentation robot writes all the manuals
6. All coordinate perfectly through the universal communication system

Time: Days instead of months
Cost: Server costs instead of salaries
Quality: No human errors, everything tested, nothing forgotten

---
Proof It Works

Simple test command shows all 5 robots working together:
✅ Backend Robot: Working and coordinating
✅ Frontend Robot: Working and coordinating
✅ DevOps Robot: Working and coordinating
✅ QA Robot: Working and coordinating
✅ Documentation Robot: Working and coordinating

Translation: I have a fully functional AI development team that builds complete software systems automatically.

---
What's Next

Industry-specific robot factories:
- Lead Generation Factory - 12 robots that find and convert sales leads automatically
- E-commerce Factory - Robots that build and run online stores
- Marketing Factory - Robots that create and manage marketing campaigns

The vision: Never manually build software again. Just tell the robots what you want, they coordinate and build it.

****************************************************************************
****************************************************************************
****************************************************************************

What I Mean by "Lead Gen Factory"

Instead of manually building 12 individual lead generation agents, we create a specialized Meta-Agent Factory that's designed specifically to create,
deploy, and manage lead generation agents.

Think of it like this:
- Current Meta-Agent Factory = Builds general-purpose agents (Backend, Frontend, DevOps, etc.)
- Lead Gen Meta-Agent Factory = Specialized factory that ONLY builds lead generation agents

How It Would Work:

The Lead Gen Factory Components:

1. Lead Generation PRD Parser

- Specialized for lead gen requirements
- Understands lead generation concepts: "conversion optimization", "intent signals", "enrichment APIs"
- Converts lead gen business requirements into technical specifications

2. Lead Gen Scaffold Generator

- Templates specifically for lead generation agents:
- Prospector Agent template
- Enrichment Agent template
- Messaging Optimization Agent template
- etc.
- Pre-built integrations for common lead gen APIs (Google Maps, LinkedIn, enrichment services)

3. Lead Gen Infrastructure Orchestrator


- Knows how lead gen agents need to work together
- Understands data flow: Prospector → Intelligence → Messaging → Follow-up
- Handles lead gen-specific coordination patterns

4. Lead Gen Domain Agents

- Uses our proven 5 domain agents, but configured for lead generation:
- Backend Agent → Builds lead databases, API integrations, enrichment pipelines
- Frontend Agent → Builds lead management dashboards, admin interfaces
- DevOps Agent → Handles Playwright automation, cron jobs, lead source monitoring
- QA Agent → Tests lead flows, data validation, conversion tracking
- Documentation Agent → Documents lead processes, API integrations

Why This Is Better Than Building 12 Agents Individually:

Advantages:


1. Systematic Creation: Factory creates agents following proven patterns
2. Consistent Architecture: All lead gen agents share same infrastructure
3. Easy Maintenance: Update factory → all lead gen agents get improvements
4. Rapid Expansion: Want agent #13? Factory creates it in minutes, not days
5. Built-in Coordination: Factory ensures all agents work together properly

Example Workflow:

# Create the Lead Gen Factory
node create-lead-gen-factory.js

# Use factory to create specific lead gen agents
lead-gen-factory create prospector-agent --sources="google-maps,linkedin" --industries="any"
lead-gen-factory create intelligence-agent --data-sources="clearbit,zoominfo" --enrichment-level="deep"
lead-gen-factory create messaging-optimizer --channels="email,sms" --ab-test-variants=5

# Factory automatically handles:
# - Agent coordination via UEP
# - Data flow between agents
# - Shared lead database schema
# - Common authentication/API management
# - Monitoring and observability

What the Factory Would Include:

Pre-Built Lead Gen Components:


- Lead Data Schema - Universal lead object structure
- API Integration Templates - Google Maps, LinkedIn, enrichment APIs
- Messaging Templates - Industry-specific outreach templates
- Automation Workflows - Playwright scripts for common lead sources
- Analytics Dashboards - Lead performance monitoring
- Coordination Patterns - How agents pass data to each other

Factory-Managed Services:

- Shared Lead Database - All agents write to same lead pool
- API Rate Limiting - Prevents agents from hitting API limits
- Error Handling - Centralized error monitoring and recovery
- Performance Optimization - Automatically scales agents based on lead volume
- Cost Management - Optimizes API usage across all agents

Concrete Example:

Traditional Approach (Build 12 Agents Manually):


Week 1: Build Prospector Agent from scratch
Week 2: Build Intelligence Agent from scratch
Week 3: Build Messaging Agent from scratch
...
Week 12: Finally have all agents, but they don't coordinate well

Lead Gen Factory Approach:

Week 1: Build Lead Gen Factory using our proven Meta-Agent Factory
Week 2: Factory creates Prospector + Intelligence + Messaging agents in 1 day
Week 3: Factory creates remaining 9 agents in 2 days
Week 4: All agents working together, optimizing and expanding

The Factory's Intelligence:


The factory would know things like:
- "Prospector agents always need to feed Intelligence agents"
- "Messaging agents need enrichment data to personalize outreach"
- "Performance agents need data from ALL other agents"
- "Demo optimization needs integration with existing SMS system"

Instead of you figuring out these connections, the factory handles it automatically.

Why This Is Powerful:

1. Speed: 12 agents in 1-2 weeks instead of 3 months
2. Quality: Uses proven Meta-Agent Factory patterns
3. Maintainability: One factory to update, not 12 individual agents
4. Scalability: Easy to add agent #13, #14, #15 as needs grow
5. Integration: Built-in coordination with your existing SMS system

Essentially, we're applying the same "meta" thinking to lead generation that we used to build our domain agents - create a system that creates systems.

bruh..............
 
Have you tried to do all the steps manually before you started creating all the agents? You know just for idea validation.
No, and it just ate up about 2 weeks of my time trying like mad to figure out how to make it work.

So long story short, other than the automatic dynamic demo reply tool that I made, none of my software works. Very disappointing but I have learned a lot about programming over the past two weeks.

The main issue is that I am a very beginner programmer, and while I have an idea in my head of what I want, translating that into an actual piece of working software is much easier said than done.

So as of yeserday, i have shifted my focus back to making money with just selling AI services. I had this idea where i would try to develop some super awesome software, but in reality I just spun my wheels for two weeks trying to figure out why it wasn't working.

Anyway, I just fired up the cold email campaigns and already have some calls booked.

Right now I am running 2 cold email campaigns:

1. General AI services
2. Database reactivation services

For the first one, I am basically saying, "AI can do a lot of stuff, let's get on a call and I'll ask you about your business. Based on your answers I'll put together a custom AI thing for you."

I've already gotten some interest from that. A lot of replies are coming from my 1st email follow up. My sequence is basically like this:

Email 1: Hey, I do AI automation and we can do basically anything. If you want some AI solutions to your problems, hit me up and let's talk.
Email 2: Here's a bit more info about what I do: 1) database reactivation, 2) custom chatbots, 3) automated lead gen and sales (brief desc of each)
Email 3: Humorous "last call" email about how they're ignoring me and my boss will beat me if they don't respond

The first email rarely gets a response, but the 2nd one gets quite a bit. My current sticking point now is turning them into actual calls. I've tried just sending a calendar link but people are lazy and won't do that, so now what I am doing is just trying to coordinate the call over email. Now if they're intereseted, I basically just say, "ok, we can do either 2 or 4, or if that doesn't work for you then pick a time on my calendar (link)."

The DBR stuff is similar, except it's much more of a "hard sell" so to speak, meaning that it just goes straight to the pitch and value add from the first email.

Both campaigns are heavily personalized with merge tags and some AI rewriting as well. Good stuff. I honestly may remove the AI rewriting as it tends to spit out very similar emails which are too vague and lame. The follow up (which I wrote) is performing much better. We'll see.

Once the cold email stuff hits full steam I think I'll be swimming in deals.
 
none of my software works
Appreciate the update! It just sounds like you've finally reached the limit of AI coding slop at this point. I've been following along what you did so far and your ideas were quite optimistic so to say. I also do not understand why you wanted to use RAG to store code from GitHub? :D If you need some more technical advice let me know!
Anyways, your AI service offer sounds quite interesting, best of luck with that, keep the updates coming
 
Last edited:
Back
Top