“Vibe coding,” Collins Dictionary’s Word of the Year for 2025, has escaped the software world. At Argonne National Laboratory, scientists at a vibe coding hackathon were speaking commands to their laptops and getting usable output in seconds. More sources are positioning it as a coming core skill for biomedical researchers.
Against that backdrop, I decided to make a video detailing how to go through the process starting from nothing. Using Claude Code and OpenAI Codex, I started with no pre-written scripts, no staged data. The first prompt told the agent to go find open-access molecular biology protocols on protocols.io, download them, read through all of them, and draft an architecture plan before writing any code. That last part, architecture first, is the lesson some vibe coding guides skip entirely, and it’s the difference between a working pipeline and a pile of disconnected scripts.
The result after one session: seven protocols downloaded, a structured extraction schema designed around what’s actually in the documents, and a working pipeline that parses messy PDFs into queryable JSON. Not perfect, the context window filled up, sub-agents went rogue, and I had to restart with a fresh agent midway through. That’s all in the video.
One thing you’ll notice: There is some chaos involved that can make narrating tricky. You’re simultaneously managing the agent’s output, planning your next prompt, watching the context window fill up, and deciding whether to let it keep going or intervene. It’s cognitively expensive in a way the hype articles never mention. It feels like herding cats toward a destination only you can see.
What’s next: a verification layer with exact metadata matching, sequence-level content diffing and regression testing against hand-reviewed golden files. Fewer vibes. More determinism. Stay tuned.
