Anna Villarreal (@annavi11arrea1) published a short, funny post on DEV Community in which she asks her favorite AI agent to roast her based on what it knows about her setup. The original article is at https://dev.to/annavi11arrea1/ultimate-ai-roast-168b and credits @trickell for the idea. The post is mostly the agent's response, reproduced verbatim, plus a prompt inviting readers to try the same with their own agent and share the results.
The prompt
The roast is triggered by a single instruction: tell the agent to roast the user based on what it knows, with permission to be harsh and use strong language. The agent, identified in the post as Opus 5.5, responds with a long, structured takedown drawn from the user's own memory file and project notes.
What the roast targets
The humor comes from specifics rather than generic insults. The agent works through the user's hardware and half-finished projects:
- A homelab with a Dell PowerEdge, a 48-port Juniper switch, a Jetson and a GPU rig, all used to build a voice assistant called Catbot that reportedly cannot recognize the word "catbot."
- A Catbot feature described as a "3D Cyber-Feline form," and an attempt to float it transparently on the desktop.
- A trading bot that has stayed in sandbox mode since July, blocked by a SQLite datetime deprecation warning.
- Two separate Ollama instances on the same port, one on IPv4 and one on IPv6, running different models.
- An Unraid server with a failing drive in slot 4 and a RAID controller that hides drives from the OS.
- A Juniper switch named "fuzzynet," connected on COM8 and routing VLANs behind a Netgear Nighthawk, with a VLAN 666 set up just to kill it.
- A bookmarks extension called NEURAL//LINKS, wired to LinkedIn scraping, a security feed and Stripe, with the security feed still not installed.
- Local FLUX image generation, Wan 2.2 video generation, QLoRA training, a six-platform CSV auto-poster, Pinterest auto-pinning and a neon sign store with full analytics and ad configuration.
The agent's closing argument is that the user is "a professional starter of projects" whose memory file reads like "a graveyard with good documentation," and it ends by telling her to finish one thing, push the branch, take the bot off sandbox mode and replace the drive.
Why it resonates
The post collected 53 reactions on DEV. Its appeal is less about the roast mechanics than about how recognizable the underlying pattern is: capable developers with elaborate infrastructure and a long tail of unfinished side projects. The agent's ability to assemble that picture from stored context is itself the interesting part, and the post's only ask is that readers repeat the experiment with their own agent.
The article does not describe the agent's configuration or how the memory file was built, so the setup behind the roast is not covered.