The "Yes-Man" Problem
We all know the struggle. You spend weekend building a complex AI agent. You stare at the code for so long that you lose all objectivity.
If you ask ChatGPT, "Is this good?", it usually replies with excessive politeness: "Great job! This architecture is robust and scalable!"
I didn't want a compliment. I wanted a Code Review. Actually, I wanted a Roast.
The Experiment: Automated Red Teaming
I recently built an Autonomous Research Agent using n8n, OpenAI and Tavily (I wrote a full technical tutorial on how to build it here- https://dev.to/practicalaiguy_ba30448492/i-built-an-ai-research-agent-to-cure-my-doomscrolling-addiction-42mi).
But before deploying it, I wanted to stress-test the logic. Since I didn't have a senior engineer handy to tear my code apart, I decided to build a synthetic one using Google NotebookLM.
The Setup
NotebookLM's "Audio Overview" feature is usually used for polite podcast summaries. But I discovered that if you use the Custom Instructions feature, you can "jailbreak" the hosts into becoming hostile personas.
I uploaded a video walkthrough of my n8n workflow and gave them this specific instruction:
The "Roast" Prompt
If you want to try this on your own projects, here is the exact prompt I used:
Character: Act as two jaded, cynical tech reviewers.
Tone:
Be highly critical.
Mock the complexity of the "no-code" graph.
Use "hacker" slang (e.g., "spaghetti code", "wrapper").
Narrative Arc:
Start: Roast the project as just another "OpenAI wrapper."
Middle: Notice the specific engineering details (specifically The Researcher - Tavily API to summarize news from multiple sources).
End: Grudgingly admit that the architecture is actually valid and solid.
The Result 💀
I expected a funny 30-second clip. What I got was a surprisingly accurate audit of my architecture.
The AI hosts:
Mocked my "Visual Spaghetti": They correctly identified the chaos of my n8n canvas. They mocked, its one among the hundreds out there, but they regretted that comment later.
Validated the Logic: By "roasting" it, they proved they actually understood why I built it that way.
Found Value : Towards the end of the video, they found out additional value and use cases (which I never thought in the first place), and admit the tool is a highly customizable prototype.
Watch the Roast
You can hear the full audio (and see the agent breakdown) here:
Why This is Actually Useful
Aside from the emotional damage, this is a legitimate productivity hack.
When you are working solo, you live in an echo chamber. Forcing an LLM to adopt a hostile persona ("The Angry Senior Dev", "The Confused User", "The Security Auditor") breaks the "Yes-Man" loop.
It forces the model to look for flaws instead of strengths.
Try it out: Upload your own documentation or code video to NotebookLM, paste the prompt above, and see if your ego survives.
Read the Build Guide: If you want to see the actual code and nodes that the AI was roasting, check out my Technical Build Tutorial here - https://dev.to/practicalaiguy_ba30448492/i-built-an-ai-research-agent-to-cure-my-doomscrolling-addiction-42mi.
Top comments (0)