Most product tours show you features. This one shows a build.
We recorded the whole path in, from the sign-up page to a system that works, and cut it into a single crash course. Sixteen minutes, no cutaways at the hard parts. Everything in it is real: real Google accounts, a real form submission, real internet research, and an app at the end that is reading the rows the workflow wrote a few minutes earlier.
The full Getting Started series is live today at generalinput.com/getting-started. Ten written guides, seven individual videos, and the crash course above if you would rather watch it in one sitting.
What actually gets built
The scenario is a staffing conference. It is a good demo because it is small enough to follow and real enough to break.
A QR code at the booth points at a Google Form. Submissions land in a Google Sheet. Every time someone fills the form in, an agentic workflow researches who they are, enriches their row in the sheet, and emails a short brief. After the conference, an app reads all of it back: headline metrics, an analytics tab, a list of every attendee, a detail page per person, and a button that drafts a follow-up email straight into Gmail.
By the end there are three durable things in the workspace: a skill that carries the writing voice, a workflow that enriches every lead as it arrives, and an app to work the results. None of them were configured in a settings page. All three were described out loud.
The seven steps
Each section below is also its own video and its own written guide, so you can go deep on one without watching the rest.
- Sign up and set up the workspace (00:40). Naming the workspace, picking a plan, and the two onboarding questions that matter more than the thirty seconds they take: what your business does, and what you personally work on. Everything an agent later writes on your behalf reads better for it. Guide
- Connect your business systems (02:48). Google Forms by one-click OAuth, then Exa for internet research by API key. Anything you connect is immediately available to every agent, workflow, and app. You do not wire it up again per automation. Guide
- Install Token Tunnel (04:01). How model selection works, what the Auto option does, and how to run your automations on the ChatGPT and Claude plans you already pay for, from a desktop app on your own machine. Your logins never leave your computer. Guide
- Build a skill (06:16). A skill built by reading the emails you have already sent, rather than by writing a style guide from scratch. It works out how you write and who you write to, then gets refined by chat with copy from the website. Guide
- Chat with your systems (09:26). The conference setup, described in one go instead of built one piece at a time. Guide
- Build the workflow (11:06). Form submission triggers research, the sheet row fills in, and a brief lands in the inbox. Then a real form submission goes through it on camera. Guide
- Build the app (13:20). Fifty-one leads in a spreadsheet turned into an interactive tool, described in the same chat thread that had been running all along, so it already had the context. Guide
Three moments worth skipping to
10:09, one message becomes three things. Rather than creating a form, then a sheet, then a QR code, the whole system gets rambled into a single message. Three artifacts come back, built across live authorized accounts rather than a copy of your data. It also pulls in the skill from the previous section without being asked twice.
11:46, the workflow tests itself. After the workflow is created it runs itself end to end with realistic data, hits a problem, fixes the workflow, republishes, and tests again. This is the part we would point at if we could only point at one thing. You are not guessing whether the first real submission will work, and you are not debugging in production with a live lead as the test case.
14:34, the app is just a conversation. The finished app has metrics, charts, a list view, detail pages, and a Gmail draft flow. It took about seven minutes to build. Restyling it at the end is not a rebuild, it is a message in the chat next to it, and the app updates live.
What you should take from it
Four things generalize past the conference demo.
Connect once, use everywhere. Integrations are authorized at the workspace level, not per automation. The Gmail connection made for the skill is the same one the app drafts through an hour later.
Context is a resource, not a preamble. The skill built at 06:16 is used by the chat at 10:09, the workflow at 11:06, and the app at 13:48. Tone and audience stay consistent because they live in one place instead of being re-explained per task.
Describing beats configuring. Every artifact here was created by talking. The output is still inspectable afterwards: workflows are plain-text readable, every variable can be swapped, and executions stream in one at a time and can be watched live.
Setup is not a separate chore. The same chat that answers questions about your data can go and create the things you are describing, because it is working against the accounts you authorized rather than a copy of them.
If you would rather read
Every step has a written version with the exact clicks, and three more guides cover ground the crash course does not: building from a prompt if you want to start from a template instead of a blank page, connecting a custom OAuth integration for services outside the 500+ in the catalog or when you want your own branding on the consent screen, and browser automation for the sites that have no API at all.
The demo system in the video is a lead tracker because a lead tracker is easy to follow in sixteen minutes. The interesting version is the one where the systems are yours and the domain knowledge is yours. That part takes about the same amount of time.