How I build
How I take a product from the customer to software, to a beta, to a release I’ve checked live. AI writes most of the code; the calls are mine.
Document uploads in 50RH were marked live, and every upload failed. Using the live app on 3 Aug 2026 showed it, and the fix took five steps, done by 4 Aug. Since then, shipped means I’ve seen it working live, not that it merged.
That is the method in small: a product isn’t done when the code is. It’s done when the people it’s for can use it, and it stays done only if someone keeps it running. The rest of this page is how I carry a product from the customer to a release, and what it takes to keep it there.
From customer to release
These are the four stages I take a product through, and no product here has been through all four yet: 50RH went live without a beta, DeHoardMe is in beta and not released, and Lantern is a prototype. Each stage below has a real example, and where my record is thin, I say so.
Customer. A product starts with people and a problem I’ve seen up close: 50RH from judging spoken estimates on job sites, Lantern from my own family. Then I write down what I don’t know. For Lantern, once the first prototype worked, I wrote down three untested guesses about who would pay and an interview plan that asks about past behavior and never pitches, and set a rule to build nothing new until those interviews say otherwise. The prototype is how I test what I wrote down, and the customer tests that matter most haven’t run yet.
Software. I build it myself, with AI writing most of the code, through the four steps further down this page.
- Prototype first. A prototype runs on seeded data until it has earned real users. Lantern is a guided demo around a made-up family, with a written bar it must clear before a real family signs in.
- Small changes. A large change goes out as small pull requests stacked on each other, so each diff holds only new work. The first screens of 50RH’s repair-job workflow went out as four, merged in order on 3 Aug 2026, and clio’s mission-control console went out the same way.
- Review I check. A second AI model from another vendor can read each change. I check each of that model’s findings against the code before I act on it: in DeHoardMe, its headline finding on a change to the payment code was wrong.
- QA. I had AI write browser tests for 50RH’s public bid, entry and invite links and a manager’s daily loop, with accessibility scans that forced real fixes, such as a page language the whole app was missing.
Beta. Testers get numbered builds and notes in plain words.
- DeHoardMe goes to its beta testers as numbered TestFlight builds. Build 15 couldn’t be made while the iOS archive on main was failing, so I fixed the archive first. Its notes correct three promises from Build 14’s notes that the app couldn’t keep.
- The thin spot is 50RH, which has no staging environment yet: a merge to main deploys straight to production.
Release. A release ends when I’ve seen it live, and the people using it can learn it.
- For 50RH’s What’s new panel I merged the backend first, checked in production that its new route answered, asking for a sign-in rather than saying “not found”, and only then merged the app.
- 50RH’s product tour walks through each stage of a job in real screenshots from the demo account.
- Video is the thin part: the only recording so far is a walkthrough of Lantern’s seeded demo, made on 30 Aug 2026.
What comes after release is the loop back to the customer, and no product here has closed it yet, so every page on this site says “built for”, never “validated with”.
Plan, audit, build, review
Inside the software stage, feature work runs through four steps. AI does most of the work inside each one, and I start every step after the plan.
- Plan. Once a week a planner reads the roadmaps of 50RH, Lantern and this site, checks each candidate against the code as it actually is, and ranks what to build next. It has run on a schedule since 27 Aug 2026, and all it writes is a plan.
- Audit. Before any code, a separate AI reviewer reads the chosen spec and asks whether it’s the right change, whether it’s complete, and what will bite during the build. It answers go, go with changes, or hold.
- Build. The change is built in its own working copy as a draft pull request, exactly as the audited brief says, with the tests the brief names.
- Review. A code review before anything merges. Changes to sign-in, database migrations and billing get a fuller look.
I don’t approve every merge by hand. A session I started may merge once the code review is clean and the change adds tests. Public copy, including every word on this site, is never merged by the agent that wrote it. And by standing rule, nothing that runs on a schedule merges anything: clio’s scheduled jobs only write reports.
Keeping it running
On 20 Aug 2026 a spending-limit block on GitHub Actions killed that night’s database backup for 50RH. The recovery steps are written down where the next person will look. Here is what keeping 50RH running looks like.
Deploys. Every merge to main builds the API, runs the database migrations as a job that blocks the deploy if it fails, brings up a revision named for the commit and moves traffic to it, seeds the demo account, runs a smoke test against production, and records the release in Sentry. One deploy at a time; rolling back means routing traffic to an earlier revision. The app and the marketing site deploy from main on Vercel, a preview per pull request. Previews run apart from production, with one gap open until staging exists.
Data. A nightly dump goes to a private bucket and is kept for 30 days; the free database plan takes no backups of its own, so this is the only recovery path. Backups are proven by restoring them. 50RH’s monthly drill restores the newest backup into a throwaway database and checks it. The scheduled drills on 1 Sep and 1 Oct 2026 passed. Point-in-time recovery waits for the first paying tenant. A permanent account delete takes a snapshot first, and doesn’t run if the upload fails.
Watching it. Sentry on the API and the app, each event tied to the commit that shipped it. A monitor on the nightly backup alerts on a failure, a hang or a missed run. A daily probe logs in as the demo user and checks the AI feature is still switched on; it checks the switch, not the quality of the drafts. Alerts fire at 80% and 100% of the AI spend cap, and the retention alert was proved by a deliberate failure on 30 Sep 2026. Not in place: an uptime check or a status page. I’m the only one on call.
Upkeep and cost. Dependency updates arrive weekly, grouped. AI spend is capped at $2 a day, and the features switch off at the cap. When Actions minutes ran near the included amount, previews became label-gated, draft pull requests were skipped and superseded runs cancelled. The cloud services scale to zero, and each deferred spend has a trigger: a paid database plan waits for the first paying tenant. The API’s settings are rebuilt from the repository on every deploy, so nothing set by hand survives a merge. Cloud auth uses no long-lived keys.
Working with others
Most of what I build goes to someone else next: testers, the people a product is for, whoever maintains it after me. In 50RH, What’s new tells property managers and owners what changed, dated by the merge that shipped it, and the feature requests that come back through the same panel are triaged from new to shipped. When contractors were confused, the cause turned out to be that managers were never told a job was waiting on them; managers are now emailed when it is, and a contractor can pass on a job. The operator checklist is written for whoever runs the product next: 93 candidate items, each checked by a second agent, reduced to about 27, ticked and dated in the same pull request that clears one. And a ticked box on any roadmap means merged and working, not merely written. The limit: the way I work with AI is mine, and it hasn’t been run by a team yet.
How it stays right
Each product’s checks run where its code lives. The counts move, so each one carries its date.
- 50RH. The test suite runs before every merge and again after it. On main, 3,139 tests passed on 5 Oct 2026.
- Lantern. The typecheck, lint, the full test suite and a production build run on every pull request. The suite had 95 tests on 30 Sep 2026.
- DeHoardMe. Backend tests run on every pull request that touches the backend, and every push to main checks that the iOS app still builds. There were 218 tests on 1 Sep 2026.
- clio. The suite runs on every pull request and every push to main. Its own docs listed 1,015 tests on 3 Oct 2026.
One check exists to catch a product saying something untrue: 50RH’s marketing tests fail if the copy says 50RH verifies, vets or guarantees a contractor, or if the unfinished sensor hardware comes back to the main navigation.
How I use AI
AI writes most of the code. In 50RH, 604 of the 761 commits on main in 2026, up to 4 Oct, carry an AI co-author line. That’s the point, not a caveat: it’s how one person runs several products.
I hand AI the weekly plan, the audit, the code and its tests, a first review, and the first draft of this site’s copy. What stays with me: what to build, the go or hold after an audit, using the thing live before I call it shipped, the words that go public, and every call where I go against the analysis.
The call
The analysis said: to save build minutes, stop running 50RH’s backend tests after each merge to main.
I closed that pull request within about a minute and restored the type check that runs after merges. That was 9 Aug 2026.
A few more of the calls, each with its date:
- In 50RH, I made the wording checks on tenant notices advise rather than block (10 Aug 2026).
- In 50RH, I won’t pay about $400 a month to test an integration until a customer on that plan brings a key (11 Aug 2026).
- In DeHoardMe, I froze new work instead of shipping the next feature its plan called for (4 Oct 2026).
I check the AI’s checkers too. Six known-stale lines were planted in clio’s docs, with the answers kept outside the repo. The monthly AI self-audit found 0 of the 6. I retired it on 1 Oct 2026 and put a plain script in its place.
Cost gets the same treatment. Every scheduled clio job has a spending cap per run, and a model with no known price is refused. The scheduled checks cost about $0.42 in the 30 days to 5 Oct 2026, read from clio’s cost ledger. That doesn’t include the flat monthly AI subscription the planning and building run on.
How this site is checked
This site is held to the same rules as the products. Its tests fail when a number has no source quoted beside it, an image has no record of where it came from, a link leads to a page the site doesn’t have, a product’s status isn’t one of the six fixed words, or a page ships any script but the analytics loader. CI runs them on every pull request. On 10 Oct 2026, all 223 of them passed. An agent may draft the words here, but every word a visitor reads is mine to approve.