Life’s gotten so convenient that “good advice” is basically everywhere—and honestly, it doesn’t always stick. In this video I took the opposite route: I built an AI therapist that only gives terrible advice, because sometimes the fastest way to learn what to do is to hear what absolutely not to do. The twist is making it sound smart while being emotionally unhinged… without crossing into stuff that’s just messed up. And yes, I spent like 15 days on this—not because the code was hard, but because I was battling the prompt the whole time. On the build side, I wired up real-time speech-to-text with Deepgram Nova-3, streaming mic audio from the browser and handling interim transcripts so the conversation feels live. Then I pipe the transcript into Claude (through a proxy) with a strict persona prompt (“terrible therapist, never break character”), store conversation history, and send the response back. Finally, I close the loop with Deepgram Aura-2 text-to-speech so the therapist talks back with surprisingly human pacing and intonation. The takeaway: the tech stack is straightforward; the real challenge is prompt design, safety boundaries, and making the experience feel natural end-to-end.