I build software
with coding agents.

“Implementation goes to Claude. Architecture decisions stay with you.”

For SEALBOX, you assigned the cryptographic core to Opus and the remaining work to Sonnet, using your own tiers based on risk. When choosing the Zenbook, you used Claude to research specifications and compare prices, but made the final purchase yourself. You also checked the links and corrected them when they did not match. Research and implementation went to the model; the brief, review and final decisions stayed with you.

“Your briefs come with architecture sketches, not vague wishes.”

The SEALBOX brief combined encryption requirements, a colour palette and interface rules in a single prompt. For the Kaltstart website, you also described exactly how the animation should behave: no elaborate custom scrolling, but a 3D flip that follows normal scroll speed. Your instructions covered both what needed to be built and how it should look and behave, giving Claude specific requirements to work from.

“You correct answers that sound plausible but are wrong.”

You caught a lookup for the wrong Haase IT company in Bochum before an application could go to the wrong address. During the Zenbook comparison, you pointed out the missing monogram design twice when the answer claimed it was only a photo angle. You kept questioning that explanation until the design change was confirmed, checking the answer against what was actually visible rather than accepting a plausible explanation.

“You check addresses, links and model numbers yourself.”

During your internship search, you checked each email address and company location against the company’s legal notice rather than relying on search results. Product details and model numbers also received that extra check. The reflection notes that you now ask whether a game offer is a key or an account before buying. Even when Claude had already supplied an answer, you checked the details yourself before acting on it.