Discussion about this post

User's avatar
Brian Mehlman's avatar

I love that AI has lowered the friction of writing code. But we train people in every serious industry for a reason. Coding is only one step in building a solution. “Vibe coding” skips design, architecture, quality assurance, and governance.

Now we have data showing the risk is real: 45–62% of AI-generated code contains security vulnerabilities; AI code introduces up to 1.7× more critical defects; production incidents are rising sharply as adoption accelerates. This isn’t about bad prompts—it’s structural. Models optimize for “works,” not “works securely.”

Build a toy app? Have fun. Build a production system? Please don’t vibe code it.

Abraham's avatar

Spent a few hours today making context files and then opening up Claude Code for the first time and putting in the prompts. Speaking as a (former?) ChatGPT devotee, these are really cool, specific, impressive results, and the quality of writing is a cut above as well. Looking forward to next week’s post about giving it instructions. Are you also gonna get into creating that general .md file that Anthropic recommends, to keep track of (and ferret out) bugs? Thanks again, keep up the amazing work… you are demystifying this in a wonderful way

2 more comments...

No posts

Ready for more?