Max Johnson, founder of AI agency briix, regularly shares practical AI guidance for business owners and founders. In a recent demonstration, he showed how he converted a previously manual content-creation workflow into an automated pipeline using Claude Code, while keeping final approval and oversight in his hands.
Start: a small knowledge base and clear context
Johnson began with a compact knowledge base containing details about himself and his business, audience information, sample scripts, and notes on his writing voice. These files provided reference material Claude Code used to evaluate topics and generate scripts aligned with his brand and style.
He emphasized the importance of explaining the process in plain language. He described this approach as “vibe coding”: clearly state what you want and let the model perform the building. Still, Johnson remained responsible for the end product: he set context, approved permissions, reviewed outputs, and answered clarifying questions from Claude Code.
Building a three-stage pipeline
Johnson split the workflow into three stages:
- Stage 1: research — have Claude Code search and gather topic ideas based on the knowledge base.
- Stage 2: scoring and selection — score those ideas and select the most relevant topic.
- Stage 3: hooks and script writing — generate three hooks and expand the strongest into a full script using the knowledge base and sample content.
He wrote prompts for each stage and ran them through Claude Code, producing a working three-part pipeline where each stage was triggered manually.
Combine stages into a single agent
Next, Johnson had Claude Code combine research, scoring, hook generation, and script writing into a unified content engine that outputs a structured JSON file. He says this was the moment the separate prompts became a true “agent.”
Add a browser-based dashboard
Johnson built a local dashboard that displays generated topics, scores, hooks, and scripts. The dashboard lets him inspect results and start new runs from a browser instead of returning to the terminal. His practical tip: be creative with the dashboard design but avoid default templates — make it look intentionally designed.
Test, extend, and remove limitations
The initial version had two limits: it produced scripts only for the top-ranked topic, and new research runs still required the terminal. Johnson asked Claude Code to generate hooks and scripts for all five selected topics and to add a button on the dashboard to start a new run. The revised system produced three scripts per topic, yielding 15 script options per run.
Practical advice
For those building similar systems, Johnson recommends starting with a repetitive task you can describe clearly. Build a working prototype with your preferred coding tools, use it, and then expand it based on what you learn.
Where to watch
Johnson demonstrated the process on the podcast Zero to Agent in 30 Minutes. The episode is available on Radar and on YouTube, Spotify, and Apple; O’Reilly members can watch live.
