This is a knowledge problem. You can fix it by pointing the model to documentation (if it exists). Otherwise the model will give you the next best guess
replyVibecoded apps will have to include MCP servers so the LLM agents can interact with them directly.
replyHuh? Both Anthropic and OpenAI have computer use tools. I frequently just tell the models to test their own work. I'm a little paranoid so I only grant access to one window and make that window VMware Workstation.
replyYou combine it with /goal. I usually set a goal like "Complete the application defined in goal.md as written. Then test it end to end autonomously using Compter Use. Record all issues discovered during testing in a to-do. Then fix the issues in the to-do. Repeat testing until no more issues are discovered."
This is where some of the readers start thinking about putting some agents to work on making Photoslop.
reply