Nine emails
Nobody had asked for anything, and there was no brief. One reply came back the next afternoon, from Dr Michael Mallory, the Program Director for Media. He named a real problem: the studio lends cameras, lenses, audio and lighting kit to students, and its technician, Daniel Chedid, had been building a booking system for it himself.
Questions before code
Before any code, we asked how borrowing works today, and whether we could stand at the counter and watch a borrow and a return. We wanted to see the job, not hear it summarised.
The two people we were talking to wanted different things: the director wanted a system built, and the technician wanted the system he had already built tested. We wrote both down side by side, with how they fitted together, before the first meeting rather than during it. In the end we did both: we took his system, tested it properly, and finished it.
A disagreement nobody names turns into rework. Written down, it takes one conversation to settle.
Saying “not yet”, in writing
At the first meeting, on 8 September, Daniel asked for three more things: native mobile apps, an AI chatbot, and a rollout to the other campuses. All three were reasonable, and none could ship first. The next day we put it in writing: they would stay out of the first version so the web app could ship. He agreed.
Once the web app was solid, each came back in a smaller form, with his agreement: an app students can install on a phone instead of native apps, a rule-based assistant instead of a chatbot, and a record for each studio instead of a rollout across campuses. The language model behind the assistant stays switched off until the university decides what data may leave it.
Deferring something is not refusing it. It is choosing the order, and writing it down so nobody remembers the meeting differently.
Building on his system
The handover, on 9 September, arrived as a zip. The whole system lived on one laptop and had never been deployed, so before changing a line we put it in a private repository.
Then we measured it: 124,101 hand-written lines, a real six-layer architecture and 481 integration tests, including one of two people writing at the same moment. That is a serious base, so we built on it rather than starting again, and on the first day fixed three of its tests so the whole suite passed.
The front end was where the work was. Most of the interface lived inside the pages, with three shared UI files between them, so we extracted a library of 30 shared components and rebuilt every page on it. The bookings page went from 3,304 lines to 44.
Measure inherited code before deciding what to do with it.
Showing instead of describing
Daniel said the inherited interface felt too crammed. Rather than describe a redesign, we built one he could click: a design brief, three rounds of mockups, then a clickable prototype of every screen. He saw the before and the after side by side, and approved the after the same day.
People judge a thing they can use far better than a thing they have to imagine.
Green tests, broken product
By the second meeting every test passed. Then we walked the exact path the studio would take on launch day, on a fresh server, the way a student would, and no student could book anything. The request page sent dates as day-month-year, and the action behind it read year-month-day. Each half was correct on its own terms. A browser check now sends a real request through that page, so the bug cannot come back.
Two more had the same shape: a race test written off as flaky that was a real bug, failing one run in six under load, and a data-retention sweep, refused by a database role we had deliberately restricted, that had quietly stopped every reminder.
Passing tests tell you the code agrees with itself. Only the client’s path, on a clean machine, tells you the product works.
Live the next morning
At the second meeting, on 22 September, Daniel asked for something he could use himself, end to end, on a real address. The hosting plan had stalled, so we changed the app to run without a server of its own: stored files behind a storage layer, the scheduler behind a protected route, migrations inside the build. We deployed it that night on free tiers. It was live at 07:29 the next morning, and hosting it costs nothing.
| At handover | 23 September | |
|---|---|---|
| Tests passing | 481 | 989 |
| Browser checks | 0 | 481 assertions |
| Shared UI components | 3 | 30 |
| Largest page (lines) | 3,304 | 44 |
| Plain-English changelog | 0 | 108 entries |
That last row matters more than it looks. Daniel reads the changelog, not the commits: one entry per change, written for the person who runs the counter.
Most of the engineering was never visible to the client. What he saw was a version he could use, the morning after he asked for it.