Can you make a tutorial on building an auto‑dialer using cloud code, without fully decoding your SAS?
Sure, I can make that tutorial. You’ll need to connect to an infrastructural partner that provides a call room under the hood—this handles the dialing for you, so you don’t have to place calls manually. When someone picks up, you’re connected into that room. It’s a high‑level overview, but if you ask Claude Code for help, it could get you about 80 % of the way there. In practice, we worked with specialists who know how to build high‑quality dialers to finish the rest.
Your main challenge is the auto dialer itself—power dialers are complex infrastructure, not just a simple front‑end app. For example, Clearvo is far more infrastructure than a basic app that triggers three dials on Twilio. For prompts and tools we rely heavily on Claude Code, supplementing with Codex models when Claude underperforms or a new Codex model drops. GPT‑5.6 was sold a couple months ago. We’re also experimenting with Cursor in our workflow; it’s just tagging it like Claude, and I find it a little bit better. Since these models change constantly, it’s not a big deal either way.
If the automation is purely procedural (no AI), you don't need to host it on cloud; you can run it on a serverless platform such as Modal or a .dev trigger, or any other server solution. Cloud Code just builds the script, which you then attach to an event like an incoming booked meeting, and it runs the procedure. If you want AI woven into the automation, use Cloud Routines—skills that run in the cloud on the client's Anthropic account. To set this up, you need a kickoff call where you explain that you'll configure a cloud routine as your infrastructure, which is easy to manage and gives you full control over your IP and infrastructure. You can gain access in three ways: have the client share their Cloud Code login, create a Teams account and add you as a user with full access, or provide their API key so you can work programmatically. I prefer full access via login or Teams because it simplifies handling missing permissions.
If the client wants a demo and you have no automation yet, build a simple MVP—it doesn't need to be perfect. You can use open-source or pre-made demos; for N8N go to its templates library, and for other cloud skills just ask the community (like mine) to get them. Do not add voice-over or heavy editing; just record a video of you talking about the demo, as I am doing. If asked whether you built it yourself, say yes—you used a mix of open-source libraries and pre-made demos, made scope-appropriate modifications, and that solves the problem.
My vibe‑coding workflow starts with a voice transcription tool where I state three things: my objective (what the app should do), the context (my business, existing tools, clients), and the constraints (what I can or can't use, like needing N8N or avoiding certain platforms). Then I define what 'done' looks—for example, being able to send a request and get a nicely packaged JSON response within 500 ms. After that I just keep working, using the voice tool to feed the AI my goal, and verify my own work until the definition of done is met. Consistency matters: feeding the same structured prompt (objective, context, constraints, definition of done) yields reliable outputs, so I stick to that pattern rather than constantly changing how I prompt.