Reading · Tech Wing Bites

Hey builders—can I call you cranes, you can pick your own nicknames?

Here’s where I’ve been: I had a stack of Fire Talk 🔥 episodes recorded and sitting on my drive but not posted.

Soooo, I made a daring plan; I sat down, edited all ten of them, and shipped them here on substack—also going out on YouTube as we speak.

That’s was my whole month—video editing, I did write some “Clint insights” section in the video post, but nothing like this.

To be honest, it drained me.

Following my own rules, I stopped, took a break, REST and didn’t build anything—seriously. 😜

Following the Builder’s Loop, after REST comes before CLARITY.

BUILD → SHIP → STUCK → REST → CLARITY → FINISH → UNSTUCK

I went one extra week out of my routine—I ran out of gas.

Therefore, I had to double up on the REST, then I got clear.

…and what came back wasn’t just build ideas, but a brand upgrade, a masterclass and new AI tools.

What's up, I'm Clint. I build private AI agents to solve real problems, and I write and talk through what it actually takes to ship and get unstuck.
New here? Start with The Psychology of Builders — the map for everything I write.

My Current Build: 🪽

Building a suite of private AI financial agents to stop revenue leakage for small businesses. I’m shipping agents that automate bookkeeping, catch billing details before they're forgotten, and drive client retention through automated, smart re-engagement nudges.

I read 2 articles from AI Meets Girl Boss, and a couple of notes, before I decided, it’s time for a brand upgrade.

My logo have a bird—it’s a blue crane and it’ has a name: Roy

Yep Roy. He’s now my alter ego and spirit animal 😂

…and I must say it was fun making.

So, if you need brand tips check her publication, she has birds too:

Now, maybe these articles can be more fun—anywhere in this article where I make weird references, or awkward em dashes, that’s me and Roy.

I’m telling you now so it doesn’t get weird later.

A bit about why I picked a blue crane.

They’re relentlessly territorial. In nesting season a blue crane will attack things that are not threatening it — sparrows, plovers, tortoises, cattle.

Burning everything he’s got, fighting something that was never coming for him—like my overthinking.

His head feathers go up right before he does it.

But I think you get the full picture why I picked him.

That was the plan for this week.

I’ve been building what I call my agent farm—a swarm of local private AI agents that run my business for Bookkeeping, Invoicing, etc.

That’s the article I owe you, that I was hoping to get your feedback on, and it’s coming next.

Here’s what stopped me.

I posted my Builder’s Loop on LinkedIn and someone left a comment about GTM.

Go To Market

I sat with Roy on that one for a while—we build, then we ship, then GTM; what is the strategy? 🤔

Because, I’ve got agents, got tools, I even got 2 years of articles + 15 podcast episodes and I could not tell you in one sentence, how any of it turns into somebody paying me—hence the pivot to local private agents to close financial gaps.

I’m a technical founder.

I don’t sell.

I build

Turns out that’s not a personality.

That’s a hole—the leak, I’ve been missing.

I’m a technical founder. I don’t sell. I build … that’s the hole

Roy: Researching Sales for Technical Founders

Soo, I did what I always do when something scares me—I went and read.

Here’s what I found:

Having no market need is the top reason startups die. Somewhere around 42%, only because nobody wanted your build.

That means that you only build what the customer needs. But, if there’s no market need, then you build for the market.

And that requires even more research and asking questions—10 to 15 conversations a month, and/or find the articles and the data that provides insights to the need.

Subconsciously, I may have blocked this when AI replied:

It’s not about your idea. It’s about the problem.

Customers want your solution to be about 60% good enough to be satisfied. Ok, 60 but what is good enough?

I have been polishing what I think is past 90% on things that makes me feel it’s good enough, but was it?

Then a16z presented to me that my technology is not soooo great, that it sells itself.

Nope!

Matt Watson article—Why Technical Founders Fail at Sales—claims 90% of your job as an entrepreneur is sales, and that building and selling actively fight each other.

Build it and they will come is crap. 💩

It’s not true, your build requires sales—even before shipping.

Another article by INSEAD—Building a Growth Engine for B2B Tech Start-Up— said most founders don’t have a growth system and they MUST evolve from an innovator into an orchestrator—especially in AI era.

Hold on, orchestrator comes back.

I kept digging, devpro article—The Hidden Gap in Most Software Startups—stated the thing that actually hurt: founders treat go-to-market as a secondary task.

You cannot sell generally, you MUST be specific.

Focus on who signs the checks.

I didn’t even know a sales team had all these fancy roles, SDR (Sales Development Representative), AE (Account Executive), and Solutions architect.

I’ve hated sales for way too long.

A good sales team is built like a good engineering team, based on systems, process, handoffs.

I’m learning sales, yo. 😁

So, that’s was the clarity, I needed at this stage, not another agent.

I need to learn how to sell, as a technical founder, without turning into somebody I’d hang up on.

Your build requires SALES!

My daughter is 20 and she’s into sales. She said she wanted to help me, so I needed to onboard her—give her context of the type of sales needed and POV: Tech Founder.

I sat down to build a learning course, that turned masterclass.

So, I wasn’t just building it for me anymore, as a result, whatever came out of this had to be something I could use + hand her and walk away.

Normally, I’d reach for Claude Code Fable 5, or drop into Codex Sol models—that’s my default.

But, I’d burned all of it—no more tokens:

You’ve hit your session limit · resets 4:20pm (America/Chicago)

All gone from building the invoicing agent—adding features I believed in, based on no data from any human being who had ever used it. 😳

Tortoises.

I was excited of the building learning materials and left me only one option.

Therefore, I opened Gemini.

Let me tell you, don’t sleep on Gemini, for real, it has some good power.

I created a prompt to say:

I don’t want to take a course, so help me build one, here’s my outline and you find the content.

Oh I remembered using Storybook.

Gemini has a thing called Storybook and I was sure this was the answer.

Build the storybook, done, that’s the course.

Sparrows.

It turn out to be a GEM, Gemini doesn’t seem to be supporting anymore—my dream of just having my content in a nice story book form was gone. I continued, the chat with painful back and forth, high-level fluff, nothing I can really use.

Then Gemini told me to use NotebookLM.

Which, I’ve used NotebookLM before, but just never thought of it in this context—duh, why?

Here’s a tool I already had, sitting right there, and a chatbot had to point at it for me.

Whatever, I took my outline over to NotebookLM and what begun as building just a course for me and my daughter’s learning became a group of learning materials to form the masterclass.

If you haven’t used NLM:

You can tell it to go find the actual sources to build from, in my case it found 10. I also ran my own local research agent alongside it to fill in gaps—turns out, I don’t need it.

Then in NLM, you can open Studio section

…and WOW, 🤩 Studio is genuinely, very impressive:

A one-hour audio podcast. Two AI voices going back and forth about my material. And you can interrupt them. Cut in, ask a question, they answer, they keep going. I was truly impressed.
A 6min video overview.
A slide deck.
Flashcards. Quizzes: pulled straight from the same sources and they were on point—very related.
And the Course.

Everything worked, flawlessly.

Then I clicked on the Course—a long string of text that would have detail content on sales for the tech founder.

NOPE—Fail.

It gave me a glorified outline.

10 sources, fancy studio and all that horsepower.

And what came out looked finished and it wasn’t.

It was hidden inventory, like Vibe Coding a scaffolded app—work that was looking done, and you don’t find out until the day you try to ship it.

That’s not NotebookLM being bad. That’s NotebookLM having no idea what done meant to me.

Neither would you, since I’d never clearly told it.

To define my done, I’d need to tell the NLM chat, not just the structure of the course, with chapters and sections but also the formate of each chapter very specific.

Then, I remembered—I have a template that I wrote it for my tech articles. I built it by hand some years ago, tuned it over dozens of pieces, and it was optimized and ready for this use case.

It says:

Define every term
Tell a story
Show before you teach
Give real details

It’s about 12 lines long, copied it and I dropped it into NotebookLM.

Best output, I got from NLM so far for writing a syllabus but still an outline—one packed with real details.

Still, wasn’t good enough, for what I wanted.

So, yep you guessed it, I took that back to Gemini and said build the full course, following these rules and outline.

Not out of the woods yet, supposed to be a hole in one but I got par 6.

Gemini would go shallow—like it couldn’t give me the details, how? it can do “deep research”

Therefore, it was time to push more, go deeper, push it some more.

Nudge, correct, nudge—the template got it started and then the steering was mine.

But somewhere in that back-and-forth I realized the nudging was the real lesson and that is where the PITM exercises came from—more on that in a sec.

5 books came out the other side.

I still tweaked every one of them to make it smooth.

Nonetheless, it came out to be a good masterclass.

On a model I don’t always use along with tools in the same ecosystem.

…that I only used because I was broke. 🤣

It was time to fully embraced being an orchestrator!

Behind the Build: This is my story—the ‘What’ and ‘Why.’ If you are a builder who wants Case Studies, AI workflows, agentic builds, code references, and the stuff that helps small teams do more without hiring a department. Technical Log.

Here’s my writing template—the 12 lines, written by hand, years ago:

Title
Chapter
Index of terms — list of all terms
Section: title
  Story of the lesson
Section n: title
  The Concept
  The Problem
  The Standard
  The Fix
  ...repeat per section
Chapter Assessment (answers at the bottom)
Exercise —
  create a process that forces the learner to practice the lesson

And here’s the prompt pieces that helped built the course:

RULE 1: The “Zero Assumption” Lexicon — Never assume the reader knows industry jargon. Every Book opens with a Dictionary defining every term in plain English.
RULE 3: Immersive, High-Stakes Storytelling — Every Book opens with a Case Study. Not a summary. A narrative with a protagonist, the build-up, the specific failure, the emotional toll, the lesson.
RULE 4: Rigorous Section Architecture — Sections cannot be light bullet points. They follow this exact structure: The Concept. The Problem. The Industry Standard. The Actionable Fix.
RULE 6: The Two-Part Assessment Protocol — Multiple choice with an answer key. Then a “Punch in the Mouth” (PITM) exercise: a copy-pasteable prompt that makes the reader live-roleplay the exact scenario the Book just taught.
DO NOT generate the entire course at once. Generate ONLY Book 1. Wait for my approval.

They kinda match up—I think—with index of terms, story of the lesson, concept, problem, standards, fix, assessment with answers…..and PITM.

Nothing new got invented—maybe PITM did. 🤣

It’s now written down, so you don’t have to do it by hand.

I found out later that researchers have a name for this now for writing down prompt rules.

They call them Prompt Contracts.

I didn’t know that when I wrote mine—but now we both know.

A Prompt Contract is a structured, formal specification that treats LLM prompts like APIs by defining explicit inputs, outputs, constraints, and error behaviors to ensure consistent AI performance. Instead of relying on open-ended natural language requests (often criticized as "vibe coding"), it enforces systemic architectural rules onto AI agents and models.

I’m going to leave this in, because it’s the most useful thing I think in this article.

One of my rules said this:

RULE 5: Support your claims. Use realistic statistics, industry benchmarks, or established formulas.

Realistic—I told a machine to produce numbers that sound true.

I’ve learned that it will.

AI will happily make you feel good by providing, with total confidence, realistic answers.

I would not have catch a single “realistic” answer or statement, because I’m not a sales expert—the entire reason why I built the course.

Every other rule in that prompt pushes toward structure.

That one pushes toward plausibility, is the only rule that can lie to me.

I was about to hand this to my 20yr old daughter—oops. 😬

Here’s the fix:

RULE 5: Cite only statistics you can attribute to a named, real source. If you cannot attribute it, state the principle without a number.

That took me 4 seconds to change.

With any skill, prompt you create, you can open the file, find the hole, and patch it—something, I’m learning, this has to be done periodically as LLM models evolve.

And for us, who are not LLM Model trainers.

We can’t do that with a model but we can always edit our instructions.

Bring an outline, not a topic: “Teach me sales” gets you a glorified outline. Bring the shape you want and NotebookLM will do the rest, along with your favorite chatbot.
Let it find the sources, then go read them: NotebookLM will pull them for you. My 10 is plenty; you don’t need 50. I use text-to-speech (TTS) and listen while I read, and I write down the 10%—the insights that pops out to me. That 10% is what fuels my articles along with my experiences.
Studio is better at what it already has a format for: The audio podcast, the video overview, the deck: those are great as the format is baked in. The course builder has no baked-in format—that’s where you have to do work.
Take the good outline back to a chat model and nudge: If you’ve been chatting a while and your AI knows you, maybe you’re good, otherwise it’ll go shallow. Push it, make it go deeper, then push again. The first output is never the right answer. That back-and-forth is the job.
Make it generate one book at a time: Ask for all 5 and you get 5 outlines. Ask for one and wait, and you get a book—literally, what I had to do.
Tweak the ending yourself: Mine still needed me. Yours will too.

That’s the whole thing…

NotebookLM for sources, chat, and Studio, then Gemini—or your fav AI—for the build, and your experience for the steering from you.

Hand this to an LLM, paired with a good steering harness and the output lands close to what you expected.💥

5 books—chapters, a 1hr interruptible podcast, 6min video overview, slide deck—anime style, don’t judge me—Flashcards, Quizzes, and 10 sources.

Course Objective: A masterclass for technical founders to reframe the sales process from “persuasion” to “system debugging,” utilizing data-driven metrics, multi-agent orchestration logic, behavioral analysis, and the modern revenue stack to build a scalable, authentic revenue engine.

Before we dive into the frameworks, you must master the vocabulary your buyers and investors use daily.

SaaS (Software as a Service): A software distribution model where an application is hosted in the cloud and licensed on a subscription basis, rather than bought outright.
B2B (Business-to-Business): A business model where your company sells software to other companies (e.g., selling to a bank), rather than to individual everyday consumers (B2C).
VC (Venture Capital): Investment firms that provide funding to early-stage, high-growth startups in exchange for equity (ownership) in the company.
ARR (Annual Recurring Revenue): The predictable, recurring revenue a company expects to receive from its customers over a 12-month period. It is the lifeblood metric of a SaaS startup.
Gross Margin: The percentage of revenue you retain after paying the direct costs required to deliver your software (Cost of Goods Sold).
HITL (Human-in-the-Loop): An AI architecture where human interaction is required to verify the AI’s output, label training data, or handle complex edge cases the model cannot confidently solve.
LangGraph: A framework (built by LangChain) used by developers to build highly complex, multi-agent LLM (Large Language Model) applications that can execute cyclical, multi-step tasks.

Sarah’s hands were sweating as she sat in the glass-walled boardroom of a top-tier VC firm on Sand Hill Road. A former senior engineer at Stripe, she had spent the last eight months building “ClearWire AI,” an autonomous agent that used LangGraph to reconcile complex B2B bank wire transfers.

The product was undeniably brilliant. She had $200,000 in ARR, and her user base was growing by 20% month-over-month. She opened her laptop and asked for a $3 million Seed investment.

The lead partner leaned forward. “Your growth is great, Sarah. But I’m looking at your unit economics. Your Gross Margin is 45%. You’re not selling software; you’re selling a highly specialized consulting service wrapped in a UI. We’re going to pass.”

The room went cold. Sarah was devastated. She had assumed that because her product was technically superior, the money would follow. But in the traditional SaaS world, investors expect an 80% gross margin. You write the code once, and selling the millionth copy costs almost nothing.

Sarah, however, was running an AI company. Every time her agent reconciled a wire, she paid OpenAI for token usage. She paid a data vendor to verify routing numbers. And, crucially, she paid a team of three junior analysts to manually review every single transaction the AI flagged with a confidence score below 95%.

Sarah failed to raise capital because she tried to sell her growth while ignoring her economics. She didn’t know how to debug her own revenue engine.

The Concept The baseline gross margin for a mature, traditional SaaS business is highly predictable—usually around 80% to 85%. However, for an AI startup, these margins are frequently compressed to 50%–60%. This is known as the “AI Margin Penalty.”

The Problem If you do not explicitly account for your specialized costs, you will price your product too low to survive. AI startups suffer from “Compute Overhead,” which consists of three massive variable costs:

Compute Costs: Direct, variable payments to model providers (like OpenAI or Anthropic API calls) or the AWS/GCP cloud costs to host open-source models like Llama 3.
Data Fees: Licensing costs for third-party, proprietary datasets required for Retrieval-Augmented Generation (RAG) or continuous training.
HITL Costs: Paying human beings to supervise the AI.

The Enterprise Standard Large AI players like Scale AI have normalized the concept of HITL for enterprise safety. Big tech companies understand that early AI models require human oversight. However, they offset these costs by charging massive premium enterprise contracts (often $100k+ ARR) and ruthlessly tracking their API token usage per customer to ensure they don’t lose money on heavy users.

The Actionable Fix To sell an AI product with a 50% margin to a skeptical investor or buyer, you must reframe the conversation around Terminal Margins.

Map out exactly what your HITL team costs per transaction today.
Build a projection model showing how your AI’s confidence scores improve over time as it processes more data.
Show the math: “Today, our margin is 50% because humans review 20% of outputs. In 18 months, as the model fine-tunes on our proprietary data, human review drops to 2%, resulting in an 82% terminal margin.” You must sell the trajectory, not just the current state.

The Concept Even if your terminal margins are mapped out, you might be bleeding cash through Revenue Leakage. Revenue leakage is the unnoticed, silent loss of capital. It is fundamentally distinct from churn (a customer canceling their contract). Leakage represents revenue you legally earned from an active customer, but failed to collect.

The Problem In AI SaaS, pricing is rarely a flat flat fee; it is usually “usage-based” (charging per API call, per agent action, or per token). If your billing system is not perfectly synced with your product’s actual database, a customer might run 10,000 automated workflows, but your manual invoice only bills them for the base tier of 1,000. You are paying the compute cost for 10,000 actions, but only getting paid for 1,000. This destroys your gross margin instantly.

The Enterprise Standard Enterprise tech giants like Salesforce and Snowflake do not rely on humans to send invoices. They utilize CPQ (Configure, Price, Quote) software integrated directly into their CRM to ensure complex, usage-based billing is fully automated and error-free.

The Actionable Fix You must automate your revenue stack.

Stripe Billing: Use Stripe’s API to track usage events directly from your application’s backend.
HubSpot / Salesforce: Integrate Stripe directly into your CRM. When a customer signs a contract, the CRM should automatically trigger Stripe to begin metering usage. Removing the human from the billing process plugs the leakage.

1. Based on Sarah’s case study, why did the VC reject her ClearWire AI pitch despite strong Month-over-Month growth?

A) Her LangGraph architecture was outdated.
B) Her gross margin was 45%, leading the VC to view the company as a specialized consulting service rather than a scalable software company.
C) She did not have enough enterprise logos on her website.
D) She was targeting B2C consumers instead of B2B.

2. Which of the following best defines the “Compute Overhead” that causes the AI Margin Penalty?

A) The cost of renting physical office space for software developers.
B) Variable payments for LLM API usage, third-party data licensing, and Human-in-the-Loop (HITL) personnel.
C) The marketing budget required to outspend legacy SaaS competitors.
D) The legal fees associated with drafting SaaS contracts.

3. What is the fundamental difference between “Churn” and “Revenue Leakage”?

A) They are identical terms used by different sales departments.
B) Churn is the cost of acquiring a customer, while leakage is the cost to retain them.
C) Churn is when a customer cancels their subscription entirely; Leakage is failing to collect earned money from an active customer due to operational/billing errors.
D) Leakage only applies to hardware companies, whereas churn applies to software companies.

4. How should a technical founder defend a currently low Gross Margin (e.g., 50%) to an investor?

A) Hide the HITL costs in the marketing budget.
B) Refuse to answer and pivot the conversation entirely to user growth.
C) Explain the path to “Terminal Margins” by showing how the AI’s improving confidence scores will drastically reduce human oversight costs over time.
D) Promise to immediately fire the HITL team to save money.

(Answer Key: 1. B, 2. B, 3. C, 4. C)

Founders: Reading theory isn’t enough. Copy and paste the prompt below into ChatGPT, Claude, or Gemini to practice defending your economics live.

Copy/Paste Prompt: "Act as a deeply cynical, numbers-driven Venture Capitalist. I am a technical founder pitching an AI enterprise agent. I just asked you for a $2M seed round. My current Gross Margin is 53% because of high API compute costs and my Human-in-the-Loop review team. Grill me on my unit economics. Do not let me off the hook. Grill me. Ask me hard, specific questions about my 'Compute Overhead' and my path to a 'Terminal Margin'. If my answers are vague or purely technical without business logic, tell me you are passing on the investment. Make me earn it. Start the roleplay by saying: 'x% margins? That’s terrible for software. Walk me through exactly where that money is going. But make me ask the right question to get here.'"

Books 2–5 — the full masterclass
The 1hr audio podcast
The 6min video overview
The Desk
The flashcards and the quizzes
The 10 sources NotebookLM built from
My chapter template — with descriptions
The prompt I wish I’d started with

I’m not a salesperson—I passionately hate sales, but I was learning, and also changing my attitude.

I took this course because I built it, and I built it because I had to—and wanted to connect with my daughter.

I haven’t closed anything yet—2 demos are on the calendar and 1 agent is already adopted and running.

And if you’re a salesperson.

Go at it, genuinely, critique this—tell me the blind spots.

I know what it feels like to read something confident and thin—I just spent a whole article on it.

Next: the agent farm—my agents names are all plants. 🪴

Roy is happy at this article even though we both argued on the length.

But it’s you made it this far—with ultimate gratitude.

Thank you and…

"Learn to sell. Learn to build. If you can do both, you will be unstoppable." — Naval Ravikant

Thank you, I’m Clint. I build private AI agents to solve real problems, and I write and talk through what it actually takes to ship and get unstuck. about

I hope this article finds you well, and you’ve gain something, from taking your precious time to consume my vibe in this content.

Let’s connect on sharing build ideas, feedback or just conversations.

My Current Build: 🪽

Building a suite of private AI financial agents to stop revenue leakage for small businesses. I’m shipping agents that automate bookkeeping, catch billing details before they’re forgotten, and drive client retention through automated, smart re-engagement nudges.

If you’re a builder, founder that happens to have a technical background and also hates sales. Hopefully, this content can give you confidence in selling your product or service.

I would love to get your opinion.

I Already Had the Tool. I Just Didn’t Call It One

Comments