A child sits at the table in late-morning light, pen in fist. Tell them: “Draw a sun.” They draw a yellow circle, a few rays, maybe a face inside. Ask for the same sun ten times — or ask ten children for one — and you get ten different pictures. All recognizably suns. None the way you imagined it.
Now put a coloring book in front of the same child. The outlines are already there, the rays pre-drawn. Say: “Color it in — yellow, with a red edge.” Ten times. Now you get ten suns the way you imagined them, clean enough to hang side by side. The child didn't get smarter overnight. The difference isn't in the child. It's in the coloring book.
I draw these pages — not literally, but in every practical sense. In reality I write instructions and standards into a database, phrased neutrally enough that any task runs the same way, no matter who starts it.
Five parts, and only one of them is the model
My team consists of language models: ChatGPT, Claude. AI that writes research reports, sales material, and analyses — daily, with a consistency I could never hold by hand. And a language model is, in the decisive respect, exactly like one of those children: capable, but unreliable. Ask it the same thing ten times and you get ten variants — sometimes brilliant, sometimes off.
It draws. It produces the text.
The pre-drawn outlines the work happens inside.
What the task is executed with. If it needs a new capability (filing, research, sending), it needs another tool.
Use page 5 of the coloring book, the crayon, the colors yellow and red.
Task passed — or run it again.
The system — all five parts— is what makes the work scalable. Drawn once, used a thousand times, independent of who starts it. Together they turn randomness into something more precise, something we defined and expect: controlled randomness. Randomness with boundaries, inside our structures and standards.
Draw your own coloring book
The obvious reflex for getting better is: take a smarter child. A bigger model. And yes — a stronger model helps. On top, not instead. Hand the same coloring book and the same pen not to the child but to a highly intelligent adult: they color more cleanly, faster, with a steadier hand. But they color inside the same lines, in the same color. The picture gets tidier — not different.
This matches the research: decompose a task into clear steps with a structured instruction, and it clearly beats a naive “just do it” (Khot et al., Decomposed Prompting).
Improving a system almost never means giving the child a bigger head. It means:
- Adjusting the structure and the standard
- Adopting new tools— Google Drive, HubSpot, Apollo, Clay
- Adjusting the instructions: which standards and tools to use
- Mechanical checks and human-in-the-loop for sign-off
Produces the work, finds the accounts, writes the rows. Capable, but unreliable without structure: asked ten times, ten variants.
- Take the trigger from the trigger inbox
- Check ICP fit (profile B · anti-ICP)
- Deduplicate against the leads pool
- Prove an ICP fit score ≥ 70
- Map a buying group of ≥ 3 roles
First the mechanical check, then the human. Claude never validates itself.
Result
3–4 rows in Notion (Ainary Leads) + deal in HubSpot.
Status = Discovered · lead gate = Pending · source documented per signal
One system. Every task.
I draw these coloring book pages for your domain — research, sales, content, invoicing — so the process carries quality and scales. For founders and small teams who want to operate like a bigger one.
The better model is coming anyway, year after year, entirely without my help. Which is exactly why the work isn't worth doing on the model — it's worth doing on what remains: structure and standard.
The model draws. The coloring book decides what the picture looks like.