LampLit

Building it with AI agents from a written spec

Before most of the code existed, I decided how LampLit should behave. The first spec had 514 numbered rules in 14 areas and 69 step-by-step user flows, and I reviewed it chapter by chapter on 7 and 8 October 2026. That night, AI coding agents built the API, the phone app, the drawings and the tests from that spec. The work was split into 17 lanes, each on its own branch.

By morning the backend and app unit tests passed, and the tests that drive the whole app in a browser did not. They became a required, passing check on 20 October 2026. I added gates around the agents' work. A CI check fails any change to a decided rule unless the change log cites me, and a skeptical reviewer agent checks risky pull requests for bugs before they open. About nine in ten of the 538 commits list Claude as a co-author.

Location privacy and safety

A lit house in LampLit tells friends you're around and up for visitors, which is exactly what a controlling partner would want to know. So the app never reads the phone's location, and a light that is off looks the same as nobody home. Friends see your town or nothing. A saved address reaches only a friend you tell to come over, for that visit, or a street you chose to show it to.

Who can see a light is decided by one database query, and nothing else in the code may decide it. A live update carries no details about a light, only a signal to reload through that same query, so it can't show a light to someone outside its audience. Report and block are on every person's page, and the person is never told. Next on the list is a safety check on the notes people write on their lights.

Needing friends nearby before it is useful

A light only helps if friends who can come over are on LampLit too. A new account starts with empty streets, so sign-up is built around bringing friends along. It asks who you want to see more and says three people make a street worth checking, but it lets you go on with fewer. Invites go out as texts from your own number, with a one-time code that makes that person your friend when they join.

The iPhone version comes with a small App Clip, so an invite link opened on an iPhone without LampLit shows your friend's house and keeps the invite through the install. A helper can also set the app up for someone else, such as a parent, and hand the phone over without keeping any access. There is no way yet to include friends who won't install the app, and that idea is on the roadmap.

Getting sign-in texts delivered

Every account starts with a US mobile number, confirmed by a texted code. Carriers block texts from a toll-free number until its registration is approved, and AWS turned down LampLit's registration twice in October 2026, citing the website and the business name. The site was a JavaScript app whose plain HTML did not name the company, so I added a plain HTML home page and a line on every page that names Red Finch Advisory LLC.

On 16 October 2026 I moved sign-in codes to Twilio Verify, which sends from Twilio's own registered numbers. A server switch picks the sender, so changing it needs no app update. Amazon SES also declined production access for email, so LampLit sends email through Resend.

Store review and a new age law

Texas passed a law on app-store age checks, SB 2420. My reading of it is that apps must use the store's age data, and that a question at sign-up does not meet it on its own. So at sign-up and at every launch, LampLit asks the App Store or Google Play whether you're 18 or older, where the store has that answer. It keeps only which store confirmed you're an adult and when. It does not keep your age.

The iPhone listing's name, "LampLit: Up for Company?", has more words because the plain name Lamplit was already taken on the App Store. Before submitting, I had AI agents run mock App Store and Google Play reviews, and I fixed the clear-cut problems they found. Both stores got version 1.1.0 on 21 October 2026. Google Play published it, and the iPhone version is in App Store review.