作品集聊天 agent 用 Jev 做反垃圾和防提示注入
作者的作品集 agent 用 Jev 监控垃圾信息、路由到预设回复并防注入，跑一天流量只花 2.83 美元，无违规提示通过。
This "Chat GPTom" agent on my portfolio is built with @eve and monitored with jev for spam and for routing to prewritten responses and also for an additional layer of preventing prompt injection. I'm happy to report that after a full day of quite a bit of traffic, my spam detection, and hardening for misuse hasn't allowed any improper prompts through and has only cost me $2.83 in @vercel AI Gateway credits. For context, I have Haiku 5.5 as the main model -- it's responsible for all of the tool calls and generated responses, but I had Opus 5.5 generate a library of 79 canned responses so that not everything requires a model. I was more than a little concerned about abuse, based on what i've seen happen with work at Vercel, and I realize that I may be opening myself up for more here by posting it, but wanted to share the workflow. There's other things I have set up in addition to this that should help, but those are more nuanced. Also, I made sure to check with https://is-agentic.com/ and happy with the 95 -- though I'm close to a solid 100.