The sixth day of the journey started with numbers: 183 users (48 more since the previous stream), R$1,024 more in sales, 27 funnels running, and revenue jumping to R$2,998. But the real protagonist of the day was a fresh Anthropic launch: Claude Computer Use, which promises to control a computer on its own. Guilherme decided to test it live, having never used the tool before, throwing the agent straight at Email Hacker AI to try to send an email marketing campaign from scratch.
The test almost stalled right out of the gate: Claude managed to navigate, log in, pick a segment and product, generate the email — and got stuck exactly at confirming the send, because a native browser alert (that JavaScript confirmation box) blocked interaction with the whole page and the agent couldn't click anything but the alert itself. It was the tensest moment of the stream: instead of giving up, Guilherme decided to use Claude Code itself to rewrite the flow, swapping the alert for a custom confirmation modal, and sent Claude Computer Use to try again. On the second attempt, with the deploy already live, the send went from start to finish with no manual intervention — just Claude's mouse and keyboard.
With the path open, the stream went further: Guilherme created a second agent, the "computer explorer," with the persona of an impatient first-time user who'd never seen the tool, to simulate the whole onboarding journey on its own and report bugs the way a real user would. The exercise exposed a real list of problems — a runtime error that blew up the screen, a "free to start" CTA that in practice required an invite, screens with no visual context, buttons that didn't respond — and the agent itself gave the experience a 4.5 out of 10. From there, Guilherme fixed bug after bug live, redesigning the onboarding flow around two scenarios (a user who pastes an offer link vs. one who already has an account) and sketching a rewards system to gamify the first steps.
Midway through, Sentinel — the internal agent that monitors AI trends — flagged the launch of Claude Mobile in real time, reinforcing the thesis that the project isn't a SaaS, but a vertical email marketing agent in the mold of the Y Combinator thesis: an entire team replaced by an agent that already understands the domain. The freshly launched Discord community took on a life of its own during the stream, with members posting and getting live reviews of their own projects (from a rental-portfolio management SaaS to an app for auto-detailing shops).
The stream also solved a more mundane problem: after taking a copyright semi-strike for using third-party tracks, Guilherme generated his own AI soundtrack so he wouldn't keep putting video distribution on YouTube at risk. He closed the day after 8 hours 40 minutes of streaming, without a final deploy because he'd be traveling early the next day, leaving the public version of the product live and running, with the autonomous Computer Use send as the day's big proof of concept.