The AI Marketing Stack I Built Without Being an Engineer

I was Chief of Staff at Food Pharmer and on the founding team at Only What's Needed. My work was food labels and brand, and I'm not an engineer.

Over the last few months I built a small AI system to help with my own LinkedIn. It drafts, checks, designs and learns.

The writing and image models run free on my own laptop.

Here is what's inside, and the lesson each part taught me.

Part 1, a study before any tool

I didn't start with a tool. I started by studying what already worked for 9 health creators.

Each post was scored against its own creator's usual numbers. A big account can't win just by being big.

That study became my guide. 2 findings changed what I make:

• Posts that turned a number into something you can picture did 2.3 times better. Half the jar is sugar is easier to feel than a gram count.
• Carousels did about 1.7 times better than the creator's normal post.

The lesson for marketers is simple. Compare each post with the same creator's normal, never with the whole feed.

Part 2, rules that live in code

I have a lot of writing rules. No em dashes, no sentence starting with And, and a one-line hook.

For a while those rules lived in a document. I kept breaking them, and so did the AI.

So I moved them into 2 checkers. One runs while I draft and scores the post. The other runs on the final draft and flags anything that breaks a rule.

Some rules it enforces:

• The first line must be one sentence and fit on a phone screen.
• Words a 10-year-old wouldn't know get flagged, each with a plain swap.
• Brand names in a negative frame get flagged.
• A number with no source gets flagged.

This lesson is the most useful one in the whole setup, so here it is plainly. A rule you have to remember is one you'll break. Write it as a check.

Part 3, a learning loop that can't overreact

I note the likes, comments and reposts on my own posts. A script turns that into weights.

Those weights tell the writer which topics, formats and visuals are working for my audience.

I put 3 limits on it:

• Nothing moves off neutral until there are at least 4 posts of data.
• The top 10% of results are capped, so one viral post can't dominate.
• Every weight stays between half and 1.6 times normal.

One lucky post should never rewrite your whole strategy. The lesson is to make your learning loop slow on purpose.

Part 4, visuals where AI never draws a number

AI image models are great at mood and bad at exact digits. A wrong digit on a health post would hurt trust in me more than anything else.

So every number and table is built as a web page, then screenshotted with Playwright. That gives a sharp image every time.

For photo-style scenes I run FLUX.1-dev locally through ComfyUI. It makes the background, and the real numbers get laid on top.

The lesson is to split the job. Let the image model make the picture, and never let it write a number.

Part 5, small local models, each with one job

I use Ollama to run open models on my own graphics card. Each model does a single task:

• A drafting model writes a first version from a fact sheet. I picked it in a blind test of 5 models.
• A second model reads each draft the way a 10-year-old would, and flags anything confusing.
• A vision model reads a photo of a pack and turns the label into a fact sheet.

The label reader reads every photo twice. If the 2 reads disagree on a number, it's thrown out. A misread digit can't slip through.

I still read every draft, and I bin the ones that miss. I have thrown out more than a few.

The lesson is to test models blind on your own task. The famous name isn't always the one that writes best for you.

Part 6, a second AI whose job is to disagree

This is the part I'm proudest of.

I had a batch of sleep posts with research behind each claim. First, every claim went to one AI agent to check against the original paper.

A second agent then tried to prove the first one wrong.

23 of the 24 fact sheets needed a fix, from wrong numbers to wrong papers. A few said more than the study did.

None of those errors went live, because the check ran first.

The lesson for any marketer using AI is blunt. AI sounds sure of itself even when it's wrong. Make checking cheap, and have a second AI try to prove the first one wrong.

What you can steal without writing code

You don't need my setup to use these ideas. Here's the short version:

• Study your own niche's best posts before you copy a format.
• Judge each post against the same creator's normal.
• Turn your style guide into a checklist something can run.
• Keep your AI images away from facts and numbers.
• Test AI tools blind on your real work.
• Have a second pass that tries to prove the first one wrong.

I built all of this with AI tools after hours, one small piece at a time. The hard part was knowing the problem well, and the code came second.

Which one would help your team first, the checker or the second AI that disagrees?

Sources

Every number above comes from one of these.

• Author's own review of public posts from 9 health and food creators, 2026, each compared with that creator's usual engagement
• Author's own citation audit of 24 sleep fact sheets, August 2026 (one checking agent per claim, then one refuting agent; 23 needed corrections)
• Playwright, open-source tool by Microsoft that turns a web page into an image, playwright.dev
• FLUX.1-dev image model by Black Forest Labs; ComfyUI, open-source image workflow tool, comfy.org
• Ollama, local model runner, ollama.com