Twenty questions,
and the one that counted was the twenty-first.

James's brief fitted in a paragraph: a twenty-questions-like approach, multiple choice or a score out of ten, for the initial conditions of a software project that succeeds; comprehensive, but no more than twenty; geared to a non-technical mind; upfront that nearly anything is possible with software, though personal data and regulation get in the way; run in a person's own chat, with the result sent back to us. Health, money, scale and uptime were his four to start from.

Twenty is a ceiling, not a count. If an earlier answer settles a later question, the assistant says so in a line, records the answer and skips it, and the bar still moves; our guess is that most people answer fifteen to seventeen, and nobody has counted yet. Cutting below twenty loses a base, and adding beyond it loses the person. A score out of ten appears only where there is a real gradient, which was one question, and that one has since gone: people pick a band more honestly than they pick a number. Every question takes Other and Don't know yet, and don't know is a fine answer; it goes in the brief under "Not yet known" rather than being guessed at. Plain words, a term explained before it is used, never "impossible", and after each answer one line on what it means for the build. Only one. For a non-technical reader that line is where the twenty minutes earns itself.

The progress is one fixed line at the top of every question: the number, the group, and a bar of twenty marks that fills whether a question was asked or skipped, with a two-line recap in the person's own words after each group of four. It is the only piece of interface a chat has, and James asked for it on the first morning to keep a person in the chair. The whimsy came an hour later: the little plastic ball of the early 2000s that guessed what you were thinking in twenty questions, and was right unnervingly often. Its name is a trademark, so the copy describes it, and the frame is the toy's. Think of the software you want; after question ten the assistant says what it thinks the idea is and asks whether it is warm; after twenty it gives one sentence beginning "My guess:" that names the shape of the build and of the running. The rules forbid padding, so that is the whole of the whimsy, a frame and two guesses rather than a character.

Then the brief, in a fixed shape so it can be copied whole: the guess, the idea in the assistant's words, a care tier, the dials, what it talks to, the build, running it, the things to insist on, and what is not yet known. The questions live in one file in the repository, and a small tool stamps that file into the page, the copy anyone can download and the links; a check fails if the page falls behind, so what the page shows is exactly what the assistant reads.

The links were the morning's one surprise. Two of the chat products accept a whole prompt in the address of a new chat, which none of them documents; James clicked both with nine and a half thousand characters in the link and each opened on the first question. The third answered that the request was too large, so its link carries a short launcher that asks it to open the file from our site instead, and two more have no working way in and get a copy button. The route back is an email with a fixed subject to an address on a server we run, which was Issue No. 2: the site's content policy forbids forms, a form would need a backend, and two people read a mailbox perfectly well. From James's first message to the open pull request was forty-two minutes, and the page was live inside the hour.

The first four are the idea. What should it do, and for whom, reflected back so the first guess has something to be wrong about. Who will use it and how they get in, which decides how hard it must defend itself and the support load, and where holding passwords can be avoided. What it has to live alongside: the spreadsheets, the software already paid for and its data, and the rules the business already works under, which decide what is inherited, what must keep working meanwhile, and who else gets a say. And whether anyone has tried to build this before and what happened to it, because whatever stopped the last one is usually still there, and can be designed around once it is named.

The second four are care, and they changed shape twice on the day this was written. They began as James's domains from the first message, health and money, with personal information and where the people are, on the principle that these add care, not impossibility. By mid-morning James had seen that health and money were one question asked narrowly, and had the group recentred on it so that the set works for any industry: does yours, or this idea, have a regulator, where you are or anywhere it will be used? Only the general law; a code nobody inspects; a regulator who can ask for records, inspect or fine; a licence before it can be used at all. The slot the merge freed asks the thing the two domains were really measuring: if it got something wrong, what is the worst that happens to someone, from an inconvenience, through their money or a lot of their time, to their job, home, rights or health, to their safety. Then the kinds of personal data, which decide the privacy work, a job that recurs rather than a feature that ships, and the country, which decides which law applies and where the data may live. The brief sums these as a care tier from nought to three, defined by what each adds to the build: nothing beyond good practice; a privacy notice, retention and deletion on time; an audit trail and a record of who looked; a formal assessment and a regulator who can ask for it.

The third four are the conditions it runs in. The busiest day and whether it must feel instant, because above a few thousand people, or instant, the machines are always on and designed in from the start, and below that the simple build is the right build; "it could go viral" makes the first month slower and the tests harder. What breaks if it is down for a day, and who would notice first and who they would ring, which decides the uptime tier and names who is on call today whether they know it or not. Where people will use it, since an app in a store adds a review cycle to every change and poor signal means working offline. And what it must talk to, because each connection is a risk and a running cost: the other side changes without asking, and somebody has to notice.

The fourth four are the shape. Whether a model does part of the work inside it, to draft or to judge, and whether a person must check its output before it counts; that is Abrar's point in disguise, because if a person must check the lot, the screen where they do it is the product. The models change under you, and what was tested in March is not what runs in June. How often it will change after launch, with the note that set and forget still changes, because what it depends on moves. What must never happen, the worst one, which is the first guard built and the first thing checked again once it is live. And what they are comfortable paying each month, machines and people together, because the machines are usually the cheap part; the attention is what costs, and the question is whose.

The last four are owning it. Twelve months after launch, how will you know it was worth it, which decides what is measured from day one and whether this is a thing that gets finished or a thing that gets kept. What you can spend to get the first version live, and when you need it, because the first version is the smallest thing that starts to prove the twelve-month answer, and the money and the date say how small. Who signs it off and whose name is next to it once it is live: not who built it, but who says it is done, who gets the call when it stops, and who decides when it changes. Nobody is the commonest answer, and the most expensive. And who in the business could, today, restore a backup, read an error message, renew a certificate or tell a supplier what has gone wrong, which decides what the running needs from outside.

The division of labour was the one from the last two issues, turned around. James gave four domains and three dials; the model drafted the twenty, the rules the assistant runs them by and the brief's shape, in less time than it took to read them; James added the toy; and Abrar read the set with an assistant of his own, which came back with six findings. Nothing asks what they can spend on the build. Nothing asks who signs it off. Two questions are too long. Two overlap. The care tiers are debatable. The "Always on" list sells a bit.

All six were right about something. Spend went into the question about when the first version is needed, since money and date are the two things that decide how small it is, and "near nothing, mostly my own time" is the plainest sign that a person should self-serve. Sign-off went into the question about whose name is next to it, with a line that if that person is not the one answering, they should see the brief; that is the case that most often goes wrong with a non-technical client. The trim took the personal-data question from seven options to five, the integrations from ten to seven, the "also" clauses to one line each, and every explanation to a sentence. One overlap had already gone in the rewrite; the other was a sentence about being in the loop with whoever builds it, which belonged to the ownership questions, so it went.

The tiers had been by domain, and the review was right that a card checkout then sat in the same tier as moving money between people, and "off sick today" in the same tier as clinical records. Defining each tier by what it adds to the build fixed both, and made the number a statement about the work rather than about the domain, which is harder to argue with; when the domains themselves went later that morning, the tiers stayed. And "Always on" had listed our own practice, automated review and a preview of every change, as if it were a universal must. It is now "Insist on these", the three things a non-technical owner can check for themselves from the first day: a repository and a tracker they can see, tests that run on every change, and a deploy that is a button and can be undone. The page's last section says the same.

One number, honestly. The twenty questions came out about a fifth shorter, and the two additions and the tier definition took most of that back, so the whole prompt is a touch shorter than it was and still twelve thousand characters, twelve thousand seven hundred inside a link. And one lesson that ran the other way. Two days after the first version we stretched the same shape, a bar and lettered options and a guess, over the podcast's guest office, where a guest sits down to turn an idea into an episode, and James read it and said it would be onerous. A shape that fits a client being scoped does not fit a guest being welcomed; that one became a conversation that opens with a single question and asks only for what is missing.

The page is live, the questions are in the open, and the set will change as we learn; what the page shows is always what the assistant reads. A brief that arrives is read by two people, who reply in plain words: what we would build first, what we would leave out, and what it would cost each month, with us or without us. No sequence, no drip, no CRM. Episode 3 of the podcast is the two of us walking through the questions, one at a time, with an example from our own projects for each group of four; by the time it is recorded the set may have moved again, and the page will say so.

Of the ten questions on the front page that we said we could not answer, this touches four. What the client is buying, if the code is no longer the expensive part: the scoping before it and the running after it is this tool's answer, and it is a working answer rather than a settled one. Whether the brief is what replaces the estimate stays a question. A product manager when the specification is executable looks, from here, like twenty questions a non-technical person can answer honestly. And acceptance testing on something produced faster than a person can read it begins with two of the twenty: who signs it off, and whether a person must check what the model does before it counts.

Still open: the price, which the site promises before the pitch and has not yet published; the links, which depend on behaviour none of the chat products documents and could stop on any Tuesday, in which case the file is still there to copy; and the document that says what AgentSee is. The next issue is the one the register has held open since September, the first conversation with someone who disagrees.