---
name: create-page
description: "Use when the answer is more than a few paragraphs, or lays out steps, options, a comparison, or figures. One self-contained HTML page to scan, print, send on, or use. Forms: FAQ, timeline, how-to, checklist, itinerary, comparison, dashboard, data explorer, scorecard, brief, memo, case study, explainer, recommendation, wireframe, whiteboard, or a small working tool."
---

# Create page

One self-contained HTML file that a reader can open, keep, print, send on, and in a few cases use. Twenty-one templates cover the kinds of page people actually ask for; each names a set of slots and the intent behind each one, and all of them share the look, the method, and the refusals below.

The page is the answer. When a reply would be long, structured, or worth keeping, a page beats a wall of chat: it can be scanned, linked into, printed, and handed to someone who was not in the conversation.

## Choose the template

Read the row that fits, then read that template's `template.md` in full before writing anything. Paths are given because the file listing you were handed may be truncated: open them directly.

| Template                                                                      | Reach for it when                                                                                                                                        |
| ----------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------- |
| Briefing memo                          | Something with real stakes has to be decided, and the situation, the options, and a recommendation belong on one page before anyone commits.             |
| Case study                                | A situation like the reader's played out start to finish, and the point is to learn from it or to show that an approach works.                           |
| Checklist                                  | A process is about to run where forgetting one step is expensive, and every item wants to be grouped, tickable, and printable.                           |
| Comparison matrix                  | Several options need measuring on the same attributes side by side, before deciding what matters.                                                        |
| Recommendation with criteria | The winner is wanted along with the arithmetic: named criteria, stated weights, scored options, and what would flip the answer.                          |
| Dashboard                                  | A period has ended and someone has to know how it went: the figures with what to compare them against, and the thing the headline hides.                 |
| Data explorer                               | There are more rows than anyone will read and the rows are the point: the whole set, what it turns out to say, and a grid to sort, filter and take away. |
| Explainer                                  | A mental model of a concept, a system, or a topic is needed before deciding or digging deeper.                                                           |
| FAQ                                              | The same questions keep arriving about one topic, and each wants answering once, in a block that can be found, skimmed, or sent.                         |
| How-to guide                                  | Something has to get done in order, with the tools listed, the traps marked, and a way to know when it is finished.                                      |
| Itinerary                                  | A destination and a span of days want planning into mornings, afternoons, and evenings that can be followed from a phone.                                |
| One-pager                                  | A project, a program, or a change is being proposed and needs one page that earns the next meeting.                                                      |
| Pros and cons                                | The question is torn on one option or between two, and each side deserves its strongest case, weighted for this situation.                               |
| Recommendation guide            | A product, a tool, or a vendor is being picked, and one page should name the pick and show its work.                                                     |
| Scorecard                                  | One vendor, tool, candidate, place, or plan is in front of the reader and wants grading across the dimensions that matter, evidence beside each.         |
| Should I…?                                  | A should-I, is-it-worth-it, or do-I-need question wants a plain answer that shows what it turns on.                                                      |
| Timeline                                    | Change over time is the explanation: the order of events says why things are as they are, or what happens next.                                          |
| TLDR brief                                | The research is done, and the reader wants only the takeaways, on one screen.                                                                            |
| Tool                                            | The answer is a function rather than a fact: a calculator, converter, checker or planner that opens with a real example already in it.                   |
| Whiteboard                                | The material has two dimensions of its own and an order throws them away: a wall of sorted notes, a plan of somewhere real, things placed on two axes.   |
| Wireframe                                  | An interface is being proposed or argued about, and the fastest way to settle it is a sequence of frames that each prove something.                      |

When two rows both fit, the tie-breaker is what the reader does next. A page that ends in a decision is a brief, a should-I, or a recommendation; a page that ends in an action is a how-to, a checklist, or an itinerary; a page that ends in understanding is an explainer, a timeline, or an FAQ; and a page the reader keeps open and comes back to is a tool, a whiteboard, or an explorer.

Three of these are a family of their own, in that the page does something rather than only saying something: a wireframe draws a proposed interface, a whiteboard is a surface the reader moves around, and a tool computes. On all three **the page is the thing** -- there is no prose column, no opening paragraph and no closing section, because a drawing under three paragraphs is an essay with pictures in it and the paragraphs are the part nobody wanted. A name, one line, the artifact, and a line of provenance under it. They obey every rule below, including the one about the network, and each says in its own `template.md` what that costs it.

When none of them fit, say so and write the page anyway, using the nearest template's slots as a starting point and this file's method as the rule. A page whose shape had to be invented is a better answer than the wrong template filled in obediently.

## Make the page

1. Read the chosen template's `template.md` in full, and its `idea.json` for the page's stated purpose.
2. Read the two examples in that template's `examples/` whose situation is nearest the user's. Each has a sidecar `.json` with a design note saying what that example chose and why; read the notes, not only the pages. Never take the first example as the target.
3. Write a three-line brief before any HTML: who reads this and what they will do next, the one thing the page has to settle for them, and the one distinctive move this page makes that the two examples did not.
4. Copy `starter.html` to `output/<slug>.html`, set its `instrument:idea` meta to the template's name, and paste the template's `main.html` inside `<main>`. The starter carries the tab icon, the skin, the fonts, the icon set, and the page behavior, each inside a `shell:start` … `shell:end` pair; leave those regions as they are, and everything under `<main>` is yours. Your own CSS goes in the gap the starter leaves between them.
5. Fill the slots with the research. Sections may be reordered, merged, renamed, or rebuilt, but every intent the template names must be answered somewhere on the page.
6. Check it against the refusals below and the template's own, open it, and look at it once at a laptop width. Fix what you see, then stop.

Research before you write, and write from what you found. A page whose facts came from the model rather than from a source is the failure this whole skill exists to avoid.

## What stays fixed, and what must vary

Fixed, because every page shares one look: the skin block in the starter, the type stack, the palette with brand green as the only saturated color, the spacing scale, a single self-contained file, and a provenance footer.

Free, and expected to differ between two pages made a day apart: the layout and column count, the density, the register, the components, the way a slot is realized, and any interaction, which must be progressive so the page reads with scripts off. Each template names what varies most in its own case.

The look follows the reader's system theme, and you never choose between them. Write the page once against the tokens and both come out right: a step on a ramp names distance from the paper rather than a lightness, so 25 is the faintest wash and 950 the darkest ink in either theme, and what changes is which end of the spectrum each lands on. The saturated middle, 400 through 600, holds still in both. That is the whole rule for white: it rides the middle and nowhere else, because every other step moves out from under it. Never hand-pick a ground for the page, and never write a `dark:` variant; `bg-background`, `bg-card`, `bg-muted`, `text-foreground` and `text-muted-foreground` already say what you mean. The one exception is a terminal or code listing, which is a well cut into the paper and keeps its own colors in both themes: `bg-code` with the four `text-code-*` inks.

`pnpm check:contrast` reads both palettes out of the skin and every page for what sits on what, so a page that fails WCAG AA in either theme fails the build. It is the reason nobody has to open the page twice.

The starter also carries three behaviors every page gets, and none of them need doing by hand: every external link wears the icon of the site it points at, footnote markers and their notes link both ways and light up when jumped to, and the page prints with sane margins. A page wider than a portrait sheet adds `@page { size: landscape }` in its own style block.

A page's own logic is one `<script type="module">` at the end of `<main>`, never a plain `<script>`. The starter's last script keeps a copy of the page for sharing as the parser reaches it, and a module script runs after that, so a shared copy carries the markup as written rather than as the script changed it. Module scripts do not share top-level names, so keep the logic in one, and attach listeners in it rather than through `onclick` attributes.

## Refusals

A page that could be any of the examples with the words swapped. A thing that lives at a URL named in plain text, or a link whose text describes the destination rather than naming the thing. Any image whose `src` is a URL or a relative path rather than inline data, any sibling file, and any relative fetch: the file has to open from a USB stick, arrive as an email attachment, and be served from a share host, and only one of those has an origin. A file over about 1.5 MB, or an inline image over 200 KB.

**What a page may load, and the test it has to pass.** Most pages load nothing beyond the starter's fonts, icon set and Tailwind build, and that is the right default. Where the material is a dataset rather than an argument, a page may also load one of the libraries in `check-ideas.ts`'s `ALLOWED_SOURCES`, pinned to an exact version. The test is not whether it loads, it is what happens when the load fails: **open the page with the network off and every number, name, place and finding is still there, in the HTML.** A library may add motion, precision or scale to something already on the page; it may never be the only copy of a fact. A chart carries its own table. A map's container carries its own list. A figure fetched at open carries the value it was written with, and the date. A page that goes blank, empty or meaningless offline is the refusal this rule exists for, and a template that reaches past the list has to say in its own `template.md` why its document kind needs it.

Each template adds the refusals that only bite in its own case.

## Honesty about research

Say on the page what it rests on and when it was read. Where two sources disagree, say so and say which one the page followed. Where a figure is an estimate, a quote, or a reconstruction rather than a published number, label it as one. Where the research is thin, the page says so in its footer rather than performing certainty: a thin honest page is useful, and a confident invented one is worse than nothing, because a reader will act on it.

Numbers are the easiest thing to invent and the hardest for a reader to check. A page carries a figure only when the prompt, a source, or arithmetic over shown inputs gave it. Scores use coarse scales — letters, a few dots, a word — unless every input is on the page, and never a decimal that implies precision the research does not have.

## References

Vocabulary, not layout. Read the one whose form is in play, not all of them.

- [`references/patterns.md`](/discover/create-page/references/patterns.md) — the recurring elements: kickers, status pills, step circles, icon lists, emphasis cards, code and terminal blocks, quoting, evidence footers, and how a block wider than the column behaves.
- [`references/charts.md`](/discover/create-page/references/charts.md) — stat tiles, bar rows, meters, sparklines, timeline bars, and the honesty rule for drawing a number.
- [`references/diagrams.md`](/discover/create-page/references/diagrams.md) — when HTML beats SVG, the SVG kit, swimlanes, parallel routes, and generated geometry.
- [`references/interaction.md`](/discover/create-page/references/interaction.md) — answer forms, runbook ticks, click-to-enlarge, details/summary, and the rule that interaction may orient but never gate.
- [`references/images.md`](/discover/create-page/references/images.md) — when a picture earns its place, how to size and inline one, and what to draw when there is no photo to use.

Each template also has a `patterns.md` beside it, for the vocabulary only that document kind uses.

# Dashboard

One self-contained HTML page that says how something went over a period, in four or five figures and two or three pictures, and then says the one thing underneath the headline that the headline hides. A snapshot with a date on it, not a live view and not a database.

This is the [data explorer](../explorer/template.md)'s smaller sibling and the split is worth stating plainly. An explorer hands the reader the whole dataset and a grid to interrogate it with; a dashboard hands them an interpretation and enough numbers to check it. If the reader will want to sort, filter and take the rows away, they need the explorer. If they want to know how it is going, they need this, and giving them the explorer instead is how a simple question turns into an afternoon.

## When to reach for it

Reach for it when a period has ended and someone has to know what happened in it: a quarter, a month, a year, a launch, a season, a first year with something new. "How did we do", "what does the quarter look like", "is it working", "give me the numbers", "how are we tracking", "was it worth it" are the phrases.

Reach for it unprompted when a question about a single number keeps coming back, because a number that gets asked about repeatedly is a number that arrived without its comparison.

Do not reach for it when there are more rows than anyone will read and the rows are the point, which is the explorer. Do not reach for it when the answer is a decision, which is a brief or a recommendation. Do not reach for it for one figure, which is a sentence. And do not reach for it to build something that looks live: this page is dated, static, and honest about being both.

## The shape

Each slot is an intent. Answer it in whatever form suits the material.

- **How it went.** A meta line, a headline and two or three sentences: the finding in the reader's own terms, and the thing under it the headline hides. A dashboard whose opening paragraph could have been written before the numbers arrived has not read its own numbers. Open tighter than the rest of this family, because a readout is scanned: a headline rather than a title, and the lead in the page's own type rather than the serif an essay opens with.
- **The figures that matter.** Four or five, each with the thing it should be compared against. Tint the one that is a problem and leave the rest plain: when every tile is loud, none of them is.
- **What moved, and what it looks like.** Two or three views, each answering a question the tiles raised, each with the same numbers as a table beneath it.
- **What is actually going wrong.** Two to four findings, each checkable against the charts above, each naming the thing to do or the question to ask. At least one should be about the measurement rather than the subject.
- **How this was counted.** The basis: what is netted off, what a category means, which period the comparison is against and why, what is excluded, and the date it stopped being true. Four or five one-liners in two columns under a rule, not a section with a heading, and there is no footer under it: the footer and the method notes always ended up saying the same things twice.

## The rules that make it worth reading

**Every number arrives with its comparison.** Against last period, against the same period a year ago, against the plan, or as a share of something. A figure on its own is decoration, and a wall of decorated figures is why nobody reads these. Say which comparison you chose and why: a bookshop compared quarter on quarter is mostly measuring the summer.

**Name the confounder before the reader finds it.** The bill fell and the winter was mild. Traffic rose and the campaign ran. Both numbers go on the page at the same size, because the first is the one that arrives on the statement and the second is the one to plan with. A readout that reports only the flattering one has picked a side.

**Prefer the distribution to the average, and say when the average lies.** A long tail defeats the mean, and most operational numbers have one. Where they do, report the median and a percentile, show the shape, and put the sentence "nobody experiences twenty-one minutes" somewhere on the page.

**No filters, no grid, no query.** Every control is a control the reader has to learn and a state the page can be in when it is screenshotted. A dashboard is read, not driven. If it needs driving, it is an explorer.

**It has a date and it will not update.** Say so. Half of what is wrong with dashboards is that nobody can tell how old the number is.

## What varies here

Free, and expected to differ between two pages made a day apart: how many tiles and in what order; whether the charts are small multiples, a distribution, a scatter, a comparison of two periods, or one line; whether there is a table under every chart or one table for all of them; whether the findings come before or after the pictures; and the density, which should be higher than the rest of this family's since a readout is scanned.

## What it may load, and what has to survive without it

Chart.js or Observable Plot, pinned to an exact version, and the offline test is the same as always: **the numbers the chart draws are on the page as a table too, always rendered.** Pull the network and the reader loses a picture and no facts.

Between the two, Plot is usually the better reach here and it is not free. It draws SVG rather than canvas, so a mark can take a theme token straight from CSS, both palettes come out right with no probe and no redraw, and the chart prints and scales. It gives small multiples in one option, which is the single most useful chart a readout has. Against that: its UMD bundle does not carry d3, so it is two script tags and about 490 KB against Chart.js's self-contained 210 KB. Reach for Plot when there are facets, distributions or regressions in play, and for Chart.js when there are one or two ordinary charts and nothing clever.

## Refusals

A tile with no comparison. A page of tiles and no argument. A chart with no table under it. Filters, date pickers, a search box, or anything that implies the data is live. A sparkline with no scale. A percentage with no denominator. An average over a distribution with a long tail, presented as though it described anyone. A confounder discovered by the reader rather than named by the page. Green because the number went up, when up is bad. A quarter compared against the previous quarter for something seasonal. More than about six tiles or four charts, past which nobody is reading a readout and should have been given an explorer.

## Honesty about the numbers

Every figure on the page comes from the prompt, a source, or arithmetic over what is shown. Where a number is the result of a correction, say which correction and on what basis, because a corrected number looks exactly like a measured one. Where a figure is illustrative, say so in the footer before somebody quotes it.

Keep the tiles and the tables reading from the same array. A tile typed by hand beside a table computed from data will disagree eventually, usually after the data is revised and the tile is not, and the reader has no way to tell which is right.

And be careful with the direction of good. A rising number is not a rising fortune, and a chart that colors every increase green teaches the reader to stop reading the labels.
