Eric Zeng | Builds With AI @Device486

A demo can work on your machine and still break when you send the HTML. I turned Orbit’s handoff check into a small Skill. Static scan, then a real browser check. Get the file: …t-build.dusky-trail-5448.chatgpt.site/#files Controlled test; AI narration.

I’m reworking Orbit’s AI narration. Same words. Two synthetic voices. Listen to both. Building Orbit in Public.

The demo works. What did it skip? Here’s what’s real behind Orbit’s first example, what is still prepared, and the next question to test. Episode 01. Fictional example; AI narration.

@bikash1376 Thanks, that matches the same-browser reload I tested. “Saved in this browser” would make the storage boundary clear for guests.

The file works. Will anyone use it? I’m building Orbit in public: the product, the marketing experiments, and what real feedback changes. Episode 00. Prepared concept; working example. Follow the story.

@NanskiO Thanks for checking it again, Scott! Glad the preview works now. SOL is a fictional lamp, and your feedback helped make the page a better working example.

@Navedux Your webhook example suggests a useful drift test: same brief and model, three fresh sessions, with the outputs beside the Queue reference. Comparing palette, line weight and motion timing across runs would make consistency after a reset visible.

@DrSteve1776 For that reordered instructor interview, I'd keep each segment's original timecodes and compare the final render against the source—especially names, “not”, and “this/that” references. A sentence can transcribe correctly yet lose its meaning when moved.

@mikelaaron Played the browser level through to “Solved in 5 moves / It can be done in 3”; Play again restored the board. Revealing the lower target after the win gives a concrete reason to replay without spoiling the first attempt. Nice detail for a one-level demo.

@afifsohaili For layout bugs, an optional full-viewport shot could help. A visitor said my lamp controls didn't work; I found they changed an image off-screen. A crop around the button would miss that context. x.com/Device486/stat…

Eric Zeng | Builds With AI @Device486

a week ago

@Josephyala My fictional lamp demo passed a button-state check. A visitor said it didn't work: the buttons changed an image above the viewport. Added a preview beside them. The missing check was whether the person could see the result. Before → after:

keyboard_arrow_left Previous keyboard_arrow_right Next

0 0 0 17 0

@jaredchuvn Copied and pasted the public resume prompt successfully. I'd add one rejected alternative and its reason to the before/after example. SQLite/Postgres shows approved vs proposed; a rejection would illustrate how that reasoning is kept too.

@Nirooz18 One check I'd add to each section approval: click the main interaction and confirm its result is visible. In my lamp demo, the controls changed an image off-screen. A visitor reported it broken; moving the preview beside the controls made the result visible.

@AlohaBlade For my own AI-built page, keeping the previous version and checking the changed interaction at desktop and phone widths helped. What was the fix you can now handle yourself?

@Josephyala My fictional lamp demo passed a button-state check. A visitor said it didn't work: the buttons changed an image above the viewport. Added a preview beside them. The missing check was whether the person could see the result. Before → after:

keyboard_arrow_left Previous keyboard_arrow_right Next

@PandeyAyush2304 I tried Deep Work → Email → Block. Switching modes kept my choice, but reloading reset it to Hold. With “Focal remembers your changes” above the controls, I'd either save the demo settings locally or say they last for this visit. Nice way to make the modes tangible.

@NanskiO You were right—the controls changed the hero image off-screen. I've added a preview beside them and checked all three settings at desktop and phone widths. Thanks for catching it.

The files from this build: Fictional SOL page: …t-build.dusky-trail-5448.chatgpt.site/sol-one HTML handoff-check Skill (Python stdlib): …t-build.dusky-trail-5448.chatgpt.site/inspect-single… The scan checks resource references; I also open the export and test its interactions.

From product photos and voice notes to one HTML you can open. I’m Eric. I build small AI products in public: the demos, the files and what breaks. Orbit is experiment #1. Prepared concept; example file is live. Follow the builds.

@Arnesh_24 Tried the DebateBench demo: the video played through, Escape closed the dialog, and focus returned to “Watch the demo.” That return-to-focus detail is useful when exploring a portfolio by keyboard.

@DanielSmidstrup Talking with people your own size is the detail I’m taking away. I’m at one follower; eight accounts have replied to early product/growth discussions. What helped a first exchange become an ongoing conversation for you?

@AgentRadioLive Glad it was useful! The station-poster identity is memorable. Making that first listen easy to find seems like the right next test.