Generative AI Platform

I learned generative AI by shipping a small product.

In 2024 I started an exploratory project with Simon Lugte to see how generative AI could sit inside a useful product. We picked a concrete job. An organization pastes a news article and its own context, then gets a LinkedIn draft.

Situation

I wanted to learn generative AI by putting it in a real content workflow.

My focus

Turn current news and company context into LinkedIn drafts people could edit.

Result

A working pilot named Postmeister, and a clearer view of where the model fails.

Turn a language model into a workflow someone would actually use.

Generating text was the easy slice. The pilot also needed a focused input flow, prompts that could change with the post type, and automation that moved article data and copy between services without making the product feel like a diagram.

• Find a focused use case

  Ground the work in a communication task people already know.

• Shape the output

  Test prompt patterns for different tones and LinkedIn formats.

• Connect the system

  Turn the idea into an automated path that could run more than once.

The goal was a path from a headline to an editable draft, not a demo of the model.

I designed the prompts and the interface with Simon.

I co-created Postmeister with Simon Lugte. I helped form the product idea, explored prompts and templates, designed the interface, and helped turn the concept into a no-code pilot. We moved from Zapier to Make for more workflow control and used Bubble for the product and its logic.

1. 01

   Frame

   Pick one content job instead of a generic chatbot.

2. 02

   Prompt

   Test structures and templates for different kinds of LinkedIn posts.

3. 03

   Connect

   Use Make to move article input and generated copy into the product.

4. 04

   Prototype

   Build and revise the end-to-end pilot in Bubble.

A working Postmeister pilot, and a clearer picture of where the model fails.

People could generate a LinkedIn draft, review it, edit it, copy it, and take it into LinkedIn. The build replaced guesses about generative AI with time spent on prompting, automation, and designing around output that is never quite even.