The SDLC didn't go away
The claim
The first article asked what got skipped. Nothing did.
Analysis, design, a test plan, implementation, review. Those phases exist because software is built by people who cannot hold the whole thing in their heads at once, and no model changed that. What changed is who does the typing. A tool that appears to skip the phases has hidden them, and hidden phases are where the cost you did not budget for lives.
This film is a container starting cold and doing every phase in order. The part worth your attention is not what it produced. It is the two hundred and ten questions it refused to answer.
Reading the film
0:19 — One Word file a client sent. A table of contents and a roadmap. This is the input for everything that follows, and it was written for people to read.
1:12 — Twelve phases. Seventy work units. Five hundred and sixty hours. Added up from the units rather than guessed at the top, which is the difference between a number and an estimate. This is where that plan is earned; every other film only refers to it.
1:36 — Two hundred and ten gaps, and this is the beat to stop on. A gap is a question the plan cannot answer on its own. Sue says it plainly: most tools fill that silence with a guess, and the guess reads exactly like a decision. Every phase carries its own count, and phase one reads twenty one raised, twenty one answered.
Empty, and then filled
An open register is only half the argument, and on its own it is an excuse. The other half is what happens to those questions, and the film shows both halves of the same screen.
1:56 — Open. The register itself, each row marked OPEN against the work unit that raised it. They are real engineering questions in the customer’s own vocabulary: whether a work unit creates a placeholder execution role or waits for the one that owns IAM. Under each is an empty box reading “Type answer…”.
2:09 — Answered, and then accepted. The same rows with the answers written in, attached to the work unit that needed them. Each suggestion is labelled suggested-ai beside an Edit control, so what the machine offered and what a person accepted stay visibly different. That is a decision log a team inherits, and on week one it is worth more than the code.
On this job all two hundred and ten were answered before the build started. That is the number that matters, and it is the one to hold the product to: not how few questions it raised, but that none of them were closed by the machine on its own.
2:53 — It will not tell you something worked when it cannot see whether it did. The sentence is a card on its own, and Sue calls it the hardest promise in the product to keep. It is the subject of what it refuses, article five.
3:36 — Four seconds from one command to a running system. Not a login, not a tour. The first screen names the vocabulary it is about to use and offers four ways in, one of which is a repository you did not write.
6:05 — Take it away and it is all still there. The issues, the epics, the diagrams, the documents. Nothing to export, because none of it was ever anywhere else.
What this costs, and what it does not do
The worked example is composite. The plates come from more than one build, so the epic on screen at 1:12 and the board at 4:50 are not the same job end to end. That is honest for a demonstration of the process and it is not a claim about one customer; if you want one job held continuous without a cut, that is article three.
The five hundred and sixty hours is a plan’s estimate. The two hundred and ten gaps are unpaid work waiting for people. And a film cannot prove a phase was done well, only that it ran and left something behind you can open.
Every phase ran and left an artefact. That still does not make it a system your company can operate.
From talk to running system →See it running
unboxing in the Stellarverse index →The same film, next to every other one, with the evidence for each sitting under the film that pays for it. No price and no form on that site.