It’s a universal page layout and organization scheme for single-topic websites.
A typical instance output by the generator: worland-dog-law.online.
This page is not a description of the format. It is the format, with itself as the subject — so every claim below is demonstrated by the thing you are reading.
Below is standard on every GYRO site
A way of building a single-topic website so that one topic serves two readers at once: the newcomer who knows nothing about it, and the reader who already knows the subject well. Neither is watered down for the other, and neither has to wade through the other.
Its second idea is about footnotes. Instead of being buried at the foot of a page, tucked behind a link, or scattered wherever they fall, every reference is gathered into one column of its own — centered in the layout, given the same visual weight as the writing, and always in the same place.
They are three columns wherever they will fit as three columns. On a wide screen you see all three at once. On a phone they stack, and you read one at a time. Printed, they become one document. And where one column holds more than any column can — an entire municipal code, say — it is given its own page, as at worland-dog-law.online6. Stacked, paged or side by side, it is the same three. Usually it is a single page — no system to log into, no monthly bill, nothing that can be switched off.
Anybody who has to be believed by strangers.
The common thread is not the subject. It is that a reader might reasonably ask says who? — and deserves an answer better than a bare link.
Because one subject has to be met at more than one depth, and mixing the depths serves nobody. A customer scrolling past technical specifications stops reading. A specialist scrolling past an explanation of basic terms also stops reading.
Side by side, each picks a lane and stays in it — and can glance sideways when curious.
Every source, previewed. When column A or C cites something, the citation is a small red numeral1 — the number you just passed. That is what they are called throughout this page. Clicking one moves you to a card in the middle that tells you what the source is, who made it, what you would find there, and what it is doing in this page.
For a business this column is the sales pitch. Not "we are licensed" but a card naming the license, the issuing authority, and a link to the register where anyone can check it in ten seconds. That is a different order of credibility from a badge image.
At the top of the column, before any source, two cards name the most recent changes to the page and a third links the chronicle — the whole record of what has changed and why. A returning reader sees at once what is new. And when a numeral has brought you to a card, a Back button at its foot takes you to the exact spot you left.
They should not have to. That is the whole design.
Transparency is the property this format is built around, and it is not a sentiment — it is a set of rules, each of which costs the author something. That cost is the only reason any of them are worth anything.
For a business the same property is worth money rather than credit. The tree service sample publishes its federal employer number so that a customer can pull the state’s own record of its insurance without asking the company. No trade site does that, because it invites checking. Inviting checking is the point.
Nothing here asks to be trusted. It asks to be checked — which is a different and much cheaper thing to offer.
Because every reference on these pages names what it points at, never where it sits.
A numeral — one of the small red numbers in the text — points at the card that carries the same number in the middle column. It means the same thing whether you got here by reading from the top, by clicking a numeral in the third column, by searching, or by following somebody else’s link straight to the middle of the page. Ibid. means “the note immediately before” — and on a page read that way there is no before.
It is one property with four consequences, which is why it is worth naming rather than leaving as good manners:
Print conventions — ibid., op. cit., supra — are all relative addressing. Each saves a line by assuming a reader moving in one direction through a fixed object. Neither assumption survives a web page, which is presumably why the Chicago Manual of Style stopped recommending ibid. at its seventeenth edition, on the ground that it “risks confusing the reader”23.
The simplicity is a consequence, not the aim. It is what falls out of refusing to let any reference depend on position — and it is the property to protect first if this format is ever changed.
Because when it was first, readers clicked it immediately and never read the preview — which defeated the whole point. That was the single most important discovery in designing this.
By the time you reach the link you already know whether the trip is worth taking. Most readers stop at the preview, satisfied. That is a success, not a failure.
Links rot. Government pages move, companies fold, archives reorganize. A page that quietly turns into a row of dead ends is worse than no page.
So the links are checked2 and when one dies it is disabled, not deleted — you can still see where it pointed, the preview still says what was there, and a dated note appears at the foot of the middle column saying when the break was found.
Nothing vanishes. A reader years later can see exactly what you cited and that it was live when you cited it.
Free, without a name of your own. The hosting is free3, the security certificate is free4, and an address like yourname.github.io/site costs nothing and never lapses. About a dollar a month buys your own name instead — and that is the part you actually own, because a name you rent can be moved anywhere and a free address cannot.
The honest answer to the second half. Nobody has promised free hosting forever, and you should distrust anyone who says otherwise. What protects you is not a promise — it is that your site is a handful of plain files. If the host ever changed its terms you could move the whole thing somewhere else in an afternoon, unchanged, and keep your address. There is nothing to export, nothing to convert, and nothing held hostage.
That is the real answer to lock-in, and it is stronger than "forever" because it does not depend on anyone's goodwill.
Yes, and for a better reason than you would expect. It is not that the host has excellent security staff, though it does. It is that there is nothing there to attack.
No database, so nothing to inject. No login, so no password to steal. No forms and no server-side code, so nothing to run. No plugins, so nothing to go out of date. No cookies and no analytics, so no visitor data to lose. The entire site is text files being handed out.
The overwhelming majority of small-business sites that get hacked are running a content system with out-of-date add-ons. This has no content system and no add-ons. It cannot go out of date because there is nothing in it that has a version.
Half of that is automatic and half of it is you, and it is worth being clear which is which.
The technical half is as good as it gets by construction10. The page is small and loads almost instantly, works on a phone first, uses proper headings and structure rather than a soup of anonymous boxes, and has nothing blocking it from rendering. Those are the things a search engine measures and most sites struggle with.
The other half is your words and who links to you, and no format can supply that. Anybody promising you rankings is selling something.
Mostly yes, and the honest answer has a hole in it that is worth knowing about before somebody else finds it.
You own the domain — you are the named registrant on the public record, and ICANN’s own statement of registrant rights includes the right to move the name to a different registrar whenever you like20. That portability is what makes an address yours rather than rented.
You own the words, automatically, from the moment you write them. Nothing to file.
You do not own the hosting, and this is the hole. GitHub’s terms say plainly that you own your content and that the license you grant them is limited to running and improving the service — and equally plainly that they may “suspend or terminate your access to all or any part of the Website at any time, with or without cause, with or without notice”19. Free hosting is free because it is not a promise.
What makes that survivable is the rest of the design: the files are plain files on your own disk and the domain points wherever you tell it, so losing the host costs a Saturday rather than the site. That is the real claim — not that nothing can go wrong, but that nothing that goes wrong takes your work with it.
The full account is on Page F, including a licensing defect in these repositories that is recorded there rather than quietly fixed.
Keep it. This is not an either‑or, and anyone who tells you to delete it is wrong. A Facebook page reaches people who are not looking for you; a site of your own answers people who are. They do different jobs.
What a Facebook page cannot do is hold still.
So: both. Post on Facebook, where the people already are, and link to the page that stays put. The one you own is the one a stranger can check at nine at night without an account, and the one still there in five years whatever happens to anybody’s feed.
One honest correction, because it is the usual version of this argument: it is not true that Facebook forbids outbound links. You can post one. Whether the feed shows it to many people is a separate question nobody outside Meta can answer, so this page does not make that claim — the three above are enough, and they are checkable.
It is built for the phone first. Below a certain width the three columns become three buttons and you read one at a time; above it, all three sit side by side. Same file, no separate mobile site, nothing to keep in step.
Phone view, the button beside Save as PDF, shows this page inside a phone-sized frame at the real widths of common phones and a small tablet — so the phone layout can be checked from a desktop, without picking up a phone. On a screen that is already phone-sized it never appears, because there is nothing to show you. On this page it is there for every reader, since the layout is part of what the page is demonstrating; on a site built from the format it is for whoever builds the page, and stays hidden from everybody else.
On a long page, once you have scrolled down, a round button with an up arrow appears at the bottom right. It takes you straight back to the top.
It also prints properly, which almost nothing does anymore: all three columns unfold into one document in order, with every collapsed answer opened.
Only if you want to make the changes yourself. GitHub is where the files sit, and where the page is served from. Somebody has to put them there once — you, or whoever builds the page for you. A page that goes up and is then left alone needs nothing from anybody after that.
If you do want to change your own page, you never have to see a command line. Two free things and about twenty minutes, once:
A change then goes live like this: edit the file, open GitHub Desktop, type one line saying what you changed, click Commit, then click Push. Live in about a minute.
Two words to learn. Commit means "save this version, with a note." Push means "send it to the internet." Everything else in that application can be ignored indefinitely. Column C has the setup written out step by step.
No. One person can write it, publish it and look after it, and that is more of an advantage than it sounds. Every pair of people on a job is one more conversation to keep going: two people have one, five have ten, ten have forty-five13. The count climbs like a hockey stick, and column C draws it. Each extra person also brings their own idea of how the thing should work, and every idea has to be squared with every other. One person has none of that — no meetings, no handoffs, no waiting on anybody else before a change can go live.
One is enough, not a limit. Every change is saved with a note saying what changed and who made it, so a second person can join at any time and see how the page got to be the way it is.
Three ways, and you only need one of them.
What this format gives up. WordPress14 exists to give you an editing box: sign in, type, click Update, and the page has changed. That box is the reason for the database behind it, the account you sign in to, the plugins that extend it and the updates that never stop — and for most of what goes wrong on small sites. There is no box here, because there is nothing to sign in to. What you have instead is a file anybody can open in fifty years.
What it asks of you. If you want changes made by asking for them in plain English, you need an AI assistant and enough familiarity with one to ask for what you want and to check what comes back. That is the one skill this format assumes — and teaching it is a main point of the pilot program this format came out of11: twelve sites, four workshops open to anyone, and a few local people given a light introduction to AI, so that a subject who wants a change has somebody nearby to ask.
One person, once. Making the repository, putting the files in it, turning the hosting on, buying the domain and pointing it at the host takes about twenty minutes the first time, and column C has it written out step by step. On the sites in this family a script does the domain end of it.
The person the site is for need do none of that. On the Washakie County pilot program11 a subject chooses a name and hands over the material — the ordinance, the price list, the record of work — and signs up for nothing, installs nothing and holds no password. The site is handed over at the end, domain and all.
Yes, and this is where the format earns its keep for some trades. Each issue is dated, published the day it is written, and never revised afterward — superseded by a later issue if it turns out to be wrong, but never quietly edited.
A builder documenting a renovation, an inspector logging visits, a restorer recording a treatment, a researcher recording a day's tests5. What you end up with is not marketing. It is a dated, public, source-cited record of what was done and when, which is a materially different thing to own if anyone later disputes it.
Worth saying plainly: if that record matters to you legally, ask a lawyer how it would actually be treated. The format gives you the record; it cannot tell you what it is worth in a dispute.
Then a copy of it is somewhere else, kept by somebody who is not you.
Every page of every site built this way is sent to the Internet Archive’s Wayback Machine, which stores its own copy, at its own address, on its own hardware, with the date it was taken21. The domain can lapse, the host can change its terms, you can lose interest or die, and the pages still answer.
That is worth something on its own, but four other things come with it:
What it is not. It is not a backup — it keeps the page as a reader sees it, not the files it was built from. It does nothing whatever for search ranking, and brings no visitors. And it is close to permanent: a capture is public, and removing one means asking the Internet Archive rather than deleting a file.
The reason it belongs in this format rather than being a nicety: every page here tells its reader do not take my word for it, go and check. An archived copy is that same sentence turned on the page itself — the one link in the chain the author does not control.
Five, and deliberately unalike. Three are real sites doing real work; two are samples, built to show a trade what a page of its own would look like before anybody has to pay for one.
In use:
Samples, for the trades, the crews and the people who do not have a page yet:
All four wear SAMPLE across the masthead and carry a dashed box saying plainly that the company, the farm or the student is invented. Underneath they are built exactly as a real one would be, which is the point of showing them.
The two most recent content changes, and the record behind them.
25 September 2026 — the property underneath the others, named: every reference is absolute No “ibid.”, no “see above”. A numeral names its target rather than its position, which is why the sources can be machine-checked, why an archived copy resolves on its own, and why a quoted paragraph still refers to something. 25 September 2026 — every page now has a copy the author cannot reach All 32 pages across the four sites go to the Internet Archive, dated, by a command that verifies each capture actually landed rather than trusting the reply. What that buys, and the three things it is not, are in column A. Chronicle — what has been happening, and why The whole record in plain language, newest first: what changed, what it was for, and what is still unresolved. Kept so that anyone returning after a few weeks can pick the thread back up without reading a commit history.Every reference to some other website, previewed. The link is always the last line — read the card first and decide whether you need the round-trip distraction of visiting the original at all.
Prior know-how is assumed here, and the items become more technical toward the bottom.
One HTML file and four small scripts. No framework, no build step, no dependencies except two web fonts, and those fall back to system faces if they fail to load. A typical page is 35–100 KB and has finished rendering before a loading spinner would have appeared. Words are drafted here first in plain text — CommonMark, the precisely specified form of Markdown7 — so they outlive whatever tool made them.
The scripts are phone-view.js, top-button.js, back-link.js and search.js. Each adds one control and what that control needs, nothing more, so a page missing one of them loses that control and still works. Wherever a site carries one, it is the identical file — changed once, in the template new sites are made from, and then copied out — so no copy drifts from the others.
Three columns wide, three buttons narrow, one document printed. The same markup serves all three — nothing is duplicated, because duplicated content is content that will drift.
Position carries meaning. A source column on the far right reads as an appendix; at the foot of the page it reads as an afterthought and goes unread. Between the plain answers and the technical detail it reads as what it is: the common evidence both other columns rest on.
It also means that whichever column you are in, a citation is one glance away rather than one long scroll.
And the card comes to you. Clicking a numeral does not send you off to find something. It brings the card up to where you are looking and marks it for a moment, so the eye lands in the right place instead of searching the column. The movement is not decoration: it is the difference between evidence arriving and evidence having to be hunted for.
Then it brings you back. A Back button appears at the foot of that card, and it returns you to the very numeral you clicked, outlined in gold for a moment, so that reading begins again at the word it stopped on. Both ends of the interruption are carried, not one: the trip out and the trip home are each a single click, and the thread the reader was holding is still in their hand.
That hunt is old. Printed notes at the foot of the page go back at least to 156815, and every one of them since has asked the same thing of a reader: leave the sentence, travel to the bottom of the page or the back of the book, find the mark again, come back if you can. Each trip costs a little of the thread the reader was holding. Four and a half centuries of that is long enough to have stopped noticing the cost.
Column B opens with two blocks built exactly alike: a heading, a caption line, then the content. Internal links, captioned The two most recent content changes., holds two cards, newest first — each a date and what changed, linked to the place on the page where it changed. Two rather than one because a single line reads as thin, and three begins to look like a log. External links, captioned Every reference to some other website., holds the source cards, one for every other site the page mentions.
The headings and captions are word for word the same on every site that carries them, and External always sits below Internal, in one column. A reader who has used one GYRO site already knows where to look on the next. On a site in page form they head the page that is column B, as on worland-dog-law.online’s Ordinance Text page6, where External links is a plain list.
Fixed order, every time, so a reader learns the shape once:
1 number and kind what sort of thing this is 2 title what it is called 3 provenance who made it, when, under what license 4 preview what you would find if you went there 5 why it is cited what it is doing in THIS page 6 the link last, and only last
Items 4 and 5 do the work, and are the two a bare footnote never supplies. Item 5 is the honest one: it forces the writer to say what the source actually supports, which is where padded citation lists fall apart.
Numerals — the small red numbers that follow a claim, each one a link to the card that supports it — belong in columns A and C, never in the header. A name in the header may still be a plain link.
Arriving at a card from a numeral adds one control at its foot, below the link: a Back button, pointing the way you came — ← Back to column A, Back → to column C. It returns you to the very numeral you clicked, outlines it in gold for a moment so the eye finds it, and then goes away. It is a way back, not something the card says; the link is still the last thing the card tells you.
The obvious implementation is wrong. A browser cannot read the status of a cross-origin request: with mode:'no-cors' the response is opaque and status is 0 whether the page is healthy, missing or on fire2. An in-page checker built that way reports false deaths, which is worse than not checking at all.
So the check runs elsewhere, on a schedule, and writes a small file the page reads at load:
links.json
{ "checked": "2026-09-10",
"dead": { "r4": "2026-08-02" },
"note": "one source unreachable since 2 August 2026" }
The page fetches that file — same origin, so the status is readable — and for anything listed dead it disables the link, strikes it through, and writes the dated note at the foot of column B. If the file is missing or unreadable the page does nothing: no file means no claim, never a false alarm.
On this site, by hand. No scheduler is wired up here yet, so links.json carries the date somebody last checked the sources themselves. The page behaves the same either way: it repeats a date a person stands behind, or it says nothing at all.
The usual pitch is that the host has good security people. True, and beside the point. The reason this is hard to compromise is that the attack surface is close to empty:
no database nothing to inject no login no credentials to steal no forms no input to poison no server code nothing to execute no plugins nothing to fall out of date no cookies no visitor data held no analytics no third party in the page
Nearly every small-business site that gets compromised is running a content system with an out-of-date add-on. There is no content system here and no add-ons. Nothing in the page has a version number, so nothing in it can become an old version.
What remains is genuine and worth naming: your account can be phished, and your domain registrar account can be taken over. Both are protected by two-factor authentication, and both are worth turning on before anything else.
The pages are plain files served from a folder3, with the certificate issued automatically4. Nobody has promised that will be free forever and the format does not depend on it being so.
The protection against lock-in is structural rather than contractual: there is no database to export, no proprietary format to convert, no theme to rebuild. Copy the folder to any other static host, repoint the domain, and the site is identical. All internal links are relative, so a change of address touches one file.
Everything a search engine asks of a page's construction10 is satisfied here without effort — one small page, no render-blocking scripts, mobile layout first, real headings, real landmarks, descriptive title and meta description, clean relative URLs, no interstitials.
What no format supplies is content worth finding and other sites linking to you. Any claim beyond the technical half should be treated as sales.
One caveat specific to this design: content hidden inside collapsed answers is in the markup and therefore indexable, but a reader arriving from a search lands on a closed accordion. Put the sentence that matters in the summary line, not only inside it.
The question behind this section is the most common objection the format meets: the subject already has a Facebook page, so why build anything? The answer is not that one is better. It is that a feed and an address are different kinds of object, and only one of them can be cited.
Ordering. A feed is sorted by recency, which means the position of a fact decays whether or not the fact does. A price list posted in spring is not retracted in autumn, it is buried, and nothing distinguishes buried from withdrawn. An address does not decay: /#hours is where the hours are, permanently, and updating them is editing one line rather than posting a correction that also sinks.
Their rule, not my tool. Facebook’s robots file asks automated clients not to fetch its pages. That is a request rather than a lock — the protocol is advisory and relies on voluntary compliance16 — but the clients that matter here are exactly the well‑behaved ones: search crawlers, web archives, and link checkers like the one described above. Four fetches were attempted from an ordinary web client while this section was written — a public page, the robots file itself, the terms of service, and the business help center — and all four came back disallowed by robots.txt. A control fetch of an unrelated site in the same session succeeded, so the refusals are the destination’s rule and not a broken tool17.
That is the concrete loss. A source card in column B previews what is at the other end of a link; it cannot preview a page no client may fetch. So a Facebook page can be linked to, but it cannot be cited in this format — not by choice, but because there is nothing to put in the card.
What is not claimed here. That Facebook demotes posts containing outbound links is widely repeated and poorly evidenced; the only sources are marketing blogs inferring it from their own reach figures. It may well be true. It is not checkable from outside, so it is not used as an argument. The three points above are, which is why they are the ones on the page.
The recommendation that follows is not migration. It is that the feed does what a feed is good at — reaching people who were not looking — and links to the address that holds still.
A single breakpoint. Above it, three columns with the unselected two dimmed rather than hidden, so the whole page is scannable; a click anywhere in a dimmed column selects it, as its button would. Below it, one column at a time with the buttons acting as tabs. Nothing is a separate mobile site, so nothing can fall out of step. Phone view, beside Save as PDF, shows the page inside a phone-sized frame at the real widths of common phones and a small tablet, so the phone layout can be checked from a desktop; it never shows on a screen that is already phone-sized. On this page it is there for every reader, because the phone layout is part of what the page demonstrates: its button carries data-public. On a site made from the format it is for whoever builds the page: its button stays hidden unless the address carries ?dev, and then stays on for the rest of that browser tab — the one thing such a page ever stores, and only in the browser of whoever asked for it; ?dev=0 turns it off.
Two buttons appear only when they are useful, and neither is printed. Scroll about a screen down and a round button with an up arrow appears at the bottom right. It takes you back to the top — smoothly, or at once if your system asks for reduced motion — and hands the keyboard to the page title, so the next Tab starts from the top. The page ends with a little blank space, so the last line of the footer can scroll clear of the button on a phone.
The other is the Back button on a source card, described above: it appears when a numeral sends you to column B, and goes away once it has brought you back. Both were first asked for on washakie-ai.online11 and then written into the format, so every new site starts with them.
The box in the navigation bar searches this page and nothing else. When the page loads, search.js walks what is actually there — every answer in column A, every card in column B, every section in column C — and builds its index out of the text it finds. Nothing is kept by hand, so the index cannot fall out of step with the page: change a sentence and the next load indexes the sentence you wrote.
It earns its place here for a reason peculiar to this format: the answers in column A are folded shut, and a browser’s own find bar will not always look inside a folded answer. A reader could search for a word that is on the page and be told it is not. A result names the column and the heading it came from; choosing one opens that column, unfolds the answer if it is folded, and marks where you landed, the way a numeral does.
No index file, no service, nothing sent anywhere: it runs in the browser, over text the browser already has, and it stores nothing. With scripts off the box never appears, which is the honest state for a control that could not work.
Body text at 17 px, line height 1.62; on a wide screen each column holds about sixty characters to the line. Figure captions are set at 22 px — larger than body text, deliberately, because a caption is often the only thing a scanning reader reads; the caption under the picture at the head of this page is one. Three accent colors only: green for structure, gold for the column keys and for marking where you are, red for citations and for whatever a click has just brought you to. A link that has died turns gray and is struck through. Dark mode redefines the nine color properties, and one label color set outside them; the numerals keep their red on a white disc in both themes.
The button calls the browser's own print dialog. Print rules unhide all three columns, drop the navigation, the floating buttons and the picture at the head of the page, and start each column on a fresh page; just before printing, a script opens every folded answer, and folds them again after — so a printed copy is a linear document in the order A, B, C, and its reader is missing nothing.
The buttons are a real tab set: role="tablist", aria-selected, panels labeled by their button, arrow-key navigation8. Answers are native <details> elements, so the browser's own find searches inside them. With JavaScript disabled entirely the page still reads: all three columns render, on a phone as on a desktop, and every answer still opens with a click. What is lost is whatever moves you around — the switching, the jump from a numeral to its card, the buttons, the striking-through of dead links — and printing then leaves the answers folded.
The Top and Back buttons carry spoken labels, since neither shows much text: Back to the top of the page, and Back to where this source is cited, in column A (or C).
Written for a reader who has never done any of it. No command line at any point.
1 Make a free GitHub account.
2 Make a new repository, named for the site. Check "Public".
A repository is a folder on their servers that keeps
every version of everything in it.
3 Install GitHub Desktop and sign in.
File > Clone repository > pick the one you just made,
and choose where it sits on your computer. You now have
an ordinary folder that is joined to GitHub.
4 Put index.html in that folder.
5 On github.com, in the repository: Settings > Pages.
Source: main branch, root folder. Save.
The site is live at username.github.io/reponame.
6 OPTIONAL, and the only step that costs money.
Buy a domain. At the registrar, point it at GitHub:
four A records on @ -> 185.199.108.153
185.199.109.153
185.199.110.153
185.199.111.153
one CNAME on www -> username.github.io.
Then add a one-line file called CNAME to the folder
containing just your domain, and publish it.
Set the DNS BEFORE publishing that file, or GitHub
shows a red warning that clears itself once the
records propagate.
7 Wait for the padlock. A few minutes after the domain
resolves a certificate is issued automatically and a
checkbox appears: "Enforce HTTPS". Check it.
8 Turn on two-factor authentication on both the GitHub
account and the registrar account. These are the only
two things anyone can actually take from you.
1 Edit index.html. 2 Open GitHub Desktop. Changes are listed on the left. 3 Type one line in Summary saying what changed. 4 Click "Commit to main". 5 Click "Push origin". 6 Live in about a minute. Add ?v= and today's date to the address when checking: the host caches for ten minutes and you will otherwise be looking at yesterday.
Two words, and no more. Commit is "save this version with a note." Push is "send it to the internet." Branches, pull requests and merges exist and can be ignored indefinitely.
There is no admin panel, and that is deliberate — an admin panel is a login, and a login is the one thing this format does not have. Page D is a form that writes text for you to paste into the file. It holds nothing, protects nothing, and can be thrown away.
So a change is made to the file, by one of three hands: the person who built the page, the owner in a text editor, or an AI asked in plain English. The third is how every site in this family is kept, and it is the one to be clear-eyed about. What it asks is not programming; it is the ability to say what should be different and to read what comes back. What it costs is a subscription to whichever assistant is used, and the habit of looking at the page before it goes live. On the pilot program this format came out of11 that help is meant to be local: the workshops are there to leave a few people in the county with a light introduction to AI and the confidence to make a change for a neighbor.
What an editor can break, and how to tell: the words sit between the tags, and the tags are what the browser reads. Delete part of one and the page usually still renders, wrongly — a paragraph swallowed into a heading, a column left half-closed. The checks under Quality assurance catch that at both widths; so does opening the page and looking at it.
Put a dot for each person at the corners of a regular polygon and join every pair: each line is a conversation the team has to keep going. Three people make a triangle, three lines; four make a square and its two diagonals, six; ten make forty-five. The count is n(n−1)/2, which grows roughly with the square of the team, so it climbs slowly at first and then steeply — a hockey stick. Fred Brooks gave the same count in The Mythical Man-Month in 1975, as part of the reason that adding people to a late software project makes it later13.
If the count of lines is the real cost, the test is whether the titles disappear. On a normal web team, a page like this one would be the work of most of these people. Each is a standard job, and each one here is answered by the format rather than by a hire:
product manager the format asks the questions UX designer three columns, fixed order visual designer type and color, already set front-end engineer one file, four small scripts back-end engineer there is no server database administrator there is no database DevOps engineer publishing is a commit systems administrator the domain script security engineer nothing versioned to patch QA engineer 66 checks, and a second reader release manager the push is the release CMS administrator there is no content system accessibility engineer built into the format SEO specialist met by construction analytics engineer nothing is collected project manager nobody to coordinate with
Set those sixteen on the drawing above and you have sixteen more corners and a hundred and twenty lines between them, before anybody has written a word for the page. That is the saving, and it is not that the work goes faster: it is that the conversations never have to happen.
Two of those titles are near-twins and both are real jobs: a product manager decides what goes on the page, a project manager keeps a team in step. The format answers the first. A team of one answers the second.
Two honest limits. This holds for a page of this kind: one subject, a few thousand words, nothing bought or logged into. It does not hold for software in general, and a hospital’s scheduling system needs every one of those people. And three jobs do not go away, because they are the work itself — reading the sources, writing the plain answer, and checking that the two match. The format does not remove that labor. It removes everything standing between it and the page.
The talk is only the visible part of the cost. Brooks’s larger point was conceptual integrity: a design holds together when a single architect, or a very few, decides what goes in and what stays out, and every added mind is another set of assumptions to reconcile. A site this size needs one person: no meetings, no handoffs, no one else’s conventions to learn. Every site in this family is built that way, by one person working with an AI, which drafts and checks while the person decides and publishes. Between people, the count is zero.
One is the fewest it needs, not the most it allows. The site lives in git, which keeps every change with its author, its date and its note, so a second person can take up the same site at any time and read how it got the way it is. If two people change the same lines at once, git stops and shows both versions, and one of them chooses.
The cost of one is the other side of the same coin: with one keeper, what that person knows can leave with them. The format answers that as far as a format can. The page explains its own workings, here in column C; git keeps every change with its note; and nothing in the page’s machinery has a version to go stale while nobody is looking after it. The words can still go out of date, and links still die — the link check above says when one has.
Optional, and not what most sites will do. Each issue is dated, published the day it is written, and superseded rather than edited — a later issue may correct an earlier one, but the earlier one stays exactly as published.
Files go /index.html for the current issue and /YYYY-MM-DD/index.html for the permanent dated copy, so a link sent today still resolves next year. The version history underneath is a second, independent record of when each word was written.
A site with more than one series gives each a folder of its own, as baconian-ciphers.online does with /biliteral/ and /word-cipher/5. The dated copies are never touched again — not even to add a button the current copy has since gained.
Sites built this way can form a family, wired upward. On an instance that sits under another site in this family, the last thing on the page, below the footer, is a card headed ↑ One level up: the name of the site above it, and one line saying what that site is, in that site’s own words. worland-dog-law.online points to washakie-ai.online11, and washakie-ai.online points here. This site is the class at the top, so it carries no such card; baconian-ciphers.online carries none either, because it links upward to a site outside this family. The map at the head of the page shows the four sites and which way their links point.
An instance adds what its own subject needs. worland-dog-law.online, three pages long, widens the same search box to cover all three, so a reader on any page finds a section on another6; baconian-ciphers.online, a running record in two series, lists every issue of both on its front page5. What proves itself on one instance is written back into the format — here, and in the template new sites are made from — so the next site starts with it.
A fourth page exists, unlinked and marked not to be indexed. It is the setup sheet: the header, the picture at the head of the page, the most-recent-change card, the source cards, the one-level-up card, and the last link-check result. Its header knows the two kinds of page — a standing page and a running series — and each block is written out as text to paste into the file.
It is not an admin panel and has no login, because there is nothing behind it to protect. It is a form that produces text you paste into a file. It is kept out of the navigation because it belongs to whoever builds the site, not to whoever reads it. Somebody who came to read about a town’s dog ordinance has no use for a form that writes HTML, and putting it in the menu would only invite the question of what it is for.
Changes are checked in two different ways before they are handed over for publishing, because each catches what the other misses.
By machine. A script drives a real browser12 over this page and its template, at desktop width (1440 px) and phone width (390 px), in light and dark, and makes 68 checks: no sideways scrolling and no script errors at either width; the most-recent-change card, which must open the passage it names; numerals, which must reach their cards and come back through the Back button, from column A and from column C; the Top button; Phone view, shown to every reader here and hidden on the template until the address carries ?dev; what printing leaves out, and that every answer opens for print and folds again after; all three columns with JavaScript off, on a phone and on a desktop; and every block Page D writes. A change is not handed over until all 68 pass. The script is kept with the format’s tools, so the checks can be run again.
By a second reader. A script checks only what someone thought to test. So larger changes are also read, claim by claim, against the code and against the sites the text describes, by a reviewer that had no part in writing them — a second AI agent, given the files and told to be skeptical. Its first reading of the 21 September update found seventeen problems, several in older text: the page said folded answers printed, and they did not; it said lines were capped near 62 characters, and on a tablet they ran past a hundred; it listed as one of this page’s failures something that could not be made to happen here. Each was fixed or reworded; a second reading confirmed the fixes and caught three more, since fixed.
The links, by hand. Whether each source is still there is checked separately, as Link health, done honestly describes: every source is fetched, and links.json records the date.
A command on the author’s own machine walks the volume where the sites live, finds each one the way the link checker does — a folder holding both a CNAME and an index.html — and turns every page it finds into the address that page is served at, using the directory form where the file is an index.html. At the time of writing that is 32 pages across four sites. New sites are picked up without anyone editing a list.
Three kinds of file are deliberately left out. Superseded drafts and anything waiting to be deleted, because archiving those makes a permanent public record of work that was withdrawn. And the claude/ folder, which by convention holds session notes rather than site content — those are published as a side effect of publishing everything else, and there is no case for making them permanent.
Anything captured in the last thirty days is skipped. The Wayback Machine is asked what it already has before anything is sent, so running the command often costs almost nothing and a page is not captured twice in a week for no reason.
Success is measured by outcome, not by the reply. Submitting a page returns a response saying the request was accepted, which is not the same as a capture existing. So after everything has been sent, the command waits and then asks the Wayback Machine, page by page, whether a capture dated today is actually there, and reports only what it can confirm. Anything unconfirmed is named rather than glossed; captures can take minutes, and the next run simply sends it again.
Two modes, and the difference is an account. Without one, each page is submitted on its own and the pages are all that is kept. With a free archive.org account, two keys go in a file beside the command, and the site’s front page can be submitted with outlinks — which captures the sources cited in column B along with the page citing them22. The command works either way and says which mode it is in before it starts. If the archive begins refusing requests, the wait between pages doubles rather than the run failing.
Nothing is committed, pushed or altered on disk; the only thing written is a dated line in a log beside the command. And because a capture is public and effectively permanent21, the command lists exactly what it is about to send and asks once before sending any of it.
Per the rule in column A, and these are real: