Machine-made text has a pacing problem
Generative models write fluently, and that is exactly the problem. A model will happily produce 900 words of even, confident, medium-length sentences — no breath, no emphasis, no place to rest. The reader scrolls, the eyes glaze, and the tab closes. Nothing was wrong with the sentences. Everything was wrong with the reading experience. This site exists in that gap: its journal at machinemadeworlds.com/blog/ publishes AI-assisted essays and guides, so the design of the article page is doing argumentative work, not decorative work — it has to make machine-made text feel considered, skimmable, and worth trusting.
My claim in one line: readers do not trust AI-assisted writing because of what it says; they trust it because of how it is staged. Typography, pacing, navigable structure, and honest contrast are the staging. What follows is the working method behind this site's own article template, with before-and-after specifics you can steal.
Give the voice a body
The first decision is typographic: body copy set in a serif, interface chrome set in a sans. On this site the article body renders at 19px with a 1.8 line height in a transitional serif stack (Iowan Old Style, Palatino, Book Antiqua, Georgia), while navigation, eyebrows, and metadata stay in a system sans — a split you can verify in the live stylesheet behind machinemadeworlds.com/. The reasoning is old and well documented: serif faces with generous line height sustain long-form reading, a point Matthew Butterick makes at length in his free online book Practical Typography, and the 65-character measure this site caps its reading column at sits inside the classic 45–75 character comfort range for continuous text.
The before state is the model default: one undifferentiated voice, everything in the same weight and rhythm, headings barely larger than body copy. The after is a two-voice page — a calm serif that says "sit down and read" and a crisp sans that says "here is where you are." Machine-made drafts almost always arrive in the before state, because models do not choose type systems. That is the editor's first job, and it is done once in the stylesheet, not once per article.
Pacing is structure, not decoration
Models write at one speed. Editors conduct. A reading experience needs fast passages (short paragraphs, lists, code) and slow passages (longer exposition, blockquotes), alternating the way this site's own archive already does: the journal at machinemadeworlds.com/blog/ mixes one-minute field notes with longer essays, so a reader learns that short means "a note" and long means "an argument." The design lesson ports directly into a single article: if every paragraph is four lines long, nothing has emphasis; if one idea gets a single-line paragraph or a pull quote, the eye lands there.
Practically, I edit machine-made drafts for three pacing moves. First, break any paragraph over roughly six lines — models stack clauses, and each stacked clause is a skimming point lost. Second, convert any sentence containing three parallel examples into a list; the list is faster to scan and easier to remember. Third, protect one slow passage per section: a blockquote or a worked example that rewards the reader who stayed. This journal's own house style demonstrates the range — compare a one-minute note with a full essay and notice how the same template stretches without breaking, which is the point of designing the template around a reading measure rather than a word count.
Headings are a contract, not decoration
The most common failure in AI-assisted drafts is heading soup: three levels of headings used interchangeably, vague labels like "Overview" and "Moving forward," no stable anchors. Headings are the one structural element that serves three audiences at once — the skimming reader, the assistive-technology user navigating by headings, and the search crawler building its picture of the page. The W3C's headings tutorial makes the accessibility case directly: headings must form a meaningful outline, because screen-reader users jump between them. And Google's guidance on creating helpful content rewards pages where the structure reflects an original, genuinely useful organization rather than keyword-shaped filler — a standard machine-made text fails by default.
This site's template enforces the contract mechanically. Every h2 in an article body gets a stable, unique anchor at publish time, and the sidebar table of contents is rendered from those headings — so if a draft ships with vague headings, the weakness is displayed in the sidebar for all to see. The 2026 SEO literature the newsroom follows says the same thing from the traffic side: long-form structured content with original perspective earns citations, and structured data plus clean page structure is how crawlers confirm what readers already feel. Before: headings written for the writer ("Some thoughts on X"). After: headings written for the navigator ("Contrast is trust," "The template, piece by piece") — each one a promise the section keeps, each one linkable.
Contrast is trust
Readers decide whether to trust a page within seconds, and contrast is a large part of that first judgment — washed-out gray text on a glaring background reads as careless, and careless pages do not get the benefit of the doubt on factual claims. The baseline is formal: WCAG 2.1 success criterion 1.4.3 requires a contrast ratio of at least 4.5:1 for normal text, with 7:1 at the enhanced AAA level. Those numbers are the floor, not the ceiling; body text should clear them comfortably in both themes.
This journal's answer is documented in its own design notes at dark-mode-design: off-white text on a deep background rather than pure white on pure black, borders instead of shadows for elevation in dark mode, and a single accent per context so links and brand signals never compete. The before state was the classic failure — a light theme inverted naively, gray text sinking into a dark background, links indistinguishable from body copy. The after treats dark mode as its own design with its own contrast budget: check the muted tones used for leads and table-of-contents links against the background, keep the body ink near-maximum contrast, and never spend low-contrast treatments on anything the reader must actually read. For machine-made text specifically, this matters more than usual: the reader already suspects the words cost nothing, so the page must show that somebody paid attention.
The template, piece by piece
This site's article template is a worked example of the whole argument. Walking top to bottom through the template behind the journal: the eyebrow line names the kind and topic before the headline, so the reader knows what contract they are entering (essay or guide, design or automation). The headline makes one promise. The lead paragraph — capped at a measure slightly narrower than the body — restates that promise in plain language; if the lead and the headline disagree, the draft is not ready. The byline carries the date and the reading time, computed from the body at roughly 220 words per minute, which means the reading-time label is a check on scope: a 1-minute label on a 2,000-word draft means the draft should have been two posts.
Below the header, the reading layout pairs a sticky table of contents with the body column — the sidebar answers "how long is this and can I jump around," the 65-character column answers "can I settle in." Before this layout, long machine-made essays were walls: no map, no exits, no re-entry points. After, every section is addressable, every anchor shareable, and the "follow the journal" link sits exactly where a finishing reader looks for it. Finally the related grid closes the loop: it prefers same-topic pieces, which is why this essay links the in-cluster companion at dark-mode-design — one link is enough to give the design shelf on the design topic page a second substantial target, and the grid does the rest automatically.
An editing checklist for machine-made drafts
Everything above compresses into a pass I run on every AI-assisted draft before it ships. One: read the headings alone — if they do not tell the story, rewrite them before touching the prose. Two: check the lead against the headline and the body against the lead; three promises, all matching. Three: break the wall — no section without a list, a quote, a code sample, or a short paragraph to change the tempo. Four: verify contrast in both themes for body, muted, and link colors against the 4.5:1 floor, and confirm the dark theme is its own design per the notes in dark-mode-design, not an inversion. Five: confirm the build-log and feed plumbing a reader never sees but always benefits from — this journal's build log records what changed on the site and why, which is the same instinct as the article template: show your structure, keep your promises, let the reader check your work.
None of this requires a new tool. It requires treating the reading experience as the product — the model supplies the words, but the typography, the pacing, the anchors, and the contrast supply the trust. Design the page as though every reader arrives skeptical, because with machine-made text, they do.