The tells field guide

Forty-two signs that an interface came out of a model on its defaults, in four families: invented information, decorative filler, reflex convergence and hollow copy. Each with why models produce it, and the one question that decides whether it stays.

All documentation

Left on its defaults, a model writes the same interface again and again: the same eyebrows, the same serif, the same three cards, the same button under three names. genjutsu calls those marks tells. This page lists all forty-two that its tells module catalogues for the web, in its four families, exactly as the module ships them.

Every entry has three parts: the marker you can see, why a model produces it, and the question that decides it. None of them is wrong in itself. A pattern is a tell when the validated thesis did not name it, and a thesis can keep any of these by listing it in its Allowed patterns: line. Nothing below says what to use instead, on purpose: a replacement would only be the next default.

Entries marked seen in our runs turned up in the pages a model wrote without genjutsu in our v4 eval runs, and the homepage quotes three of them.

Check a project yourself

No model is involved. The audit genjutsu runs at the end of its pipelines is a standard-library Python script, and it runs on its own. Install genjutsu once, then run the script from the root of a project:

With the Claude Code plugin the same script sits under the plugin install, in a folder named after the version you have: ~/.claude/plugins/cache/genjutsu/genjutsu/<version>/skills/_jutsu/design-audit/scripts/audit.py.

It runs eighteen automatic checks and prints each tell it finds with its file:line and the text, for you to hold against your thesis: a finding is not a problem until the thesis has been read. It cannot see copy held in JavaScript data, such as an array of features rendered through an expression or a content file, and the entries without a check below stay manual reads.

Invented information

The test Does this datum come from the project, or from a brief request written into the thesis? If neither, it was invented.

Automatic checks in this family: tell-invented-status, tell-locale-strip, tell-placeholder-identity, tell-fake-code, tell-round-number. What they do not cover is a manual read.

Weather or locale strip

Marker: A city code, a clock and a temperature strung together in a header or footer, such as LIS 14:23 · 18°C, or a timezone badge like GMT+1.
Why it is a tell: Models add it to make a studio feel inhabited. The city, the hour and the weather are chosen for atmosphere, not read from anything.
The question: Does the project work across time zones in a way the visitor needs to see, and does the thesis say so?

Fake build status

Marker: Release strings and sync times on a page that is not a changelog: v0.6.2-rc.1, last sync 4s ago, Build 0048, BETA.
Why it is a tell: It borrows the texture of developer tools to look shipped and alive. The numbers are made up and stay frozen forever.
The question: Which real release does this number track, and where does the page read it?

Fake heritage

Marker: A founding year worn as a badge: ESTD. 2018, Since 2016, Founded MMXIX.
Why it is a tell: An established date is the quickest way to borrow trust, so models invent one when the brief gives none.
The question: Is this the project's actual founding year, and does the visitor need it?

Fake part numbers and section codes

Marker: Codes stamped on sections, products or panels that point to no catalogue: SEC-01, MK.I, REF-204, UNIT 07.
Why it is a tell: A code makes a page look engineered, so models stamp them on to borrow the authority of a spec sheet.
The question: Which real catalogue, part list or document does this code point to, and does the thesis name that system?

Fake stock counter

Marker: A live-looking scarcity count: Reservation 412 of 800, Only 3 left, 27 people viewing.
Why it is a tell: Urgency is a conversion reflex. With no inventory behind it, the counter is a claim the page cannot keep.
The question: Which system provides this number, and does it change when stock does?

Fake photo credit

Marker: Gallery captions that credit nobody real: Plate 03 · House archive, Field study no. 12, Frame XII · 35mm under stock or placeholder images.
Why it is a tell: Captions make a placeholder image look curated. The photographer, the archive and the film stock do not exist.
The question: Who took this photograph, and has the project the right to credit them?

Placeholder names and brands

Marker: John Doe, Jane Smith, Acme, Nexus, lorem ipsum, testimonials signed by people who were never asked.
Why it is a tell: These are the first names in the model's memory for "a person" and "a company". Invented but plausible names are the same tell, only harder to spot.
The question: Is this a real client, a real quote, a real name, with permission to show it?

Round or falsely precise figures

Marker: 99.99% uptime, 10,000+ teams, 10x faster, 100% satisfaction, or the reverse disguise, 47.3% with no source anywhere on the page.
Why it is a tell: Round numbers are what a model writes when it needs a number; odd decimals are what it writes when told to avoid round ones. Neither was measured.
The question: Where was this figure measured, and could the page cite that source?

Fake product drawn in divs

Marker: A hero "screenshot" built from styled boxes: a dashboard, a terminal, a task list or a code editor that renders no real product.
Why it is a tell: Models cannot see the product, so they draw what a product usually looks like. The result shows the category, not the thing being sold.
The question: Does a real screenshot, recording or component of the product exist, and if not, does the page need a picture of it at all?

Decorative filler

The test Does this element carry information, or only texture?

Automatic checks in this family: tell-numbered-eyebrow, tell-generic-step, tell-scroll-cue, tell-dot-run, tell-glow, tell-blob, tell-perpetual-motion. What they do not cover is a manual read.

Numbered eyebrow

seen in our runs Design studio landing: 03 · Start a project over the contact section, found by audit.py.

Marker: A zero-padded index above each section title: 00 / INDEX, 001 · Capabilities, 04. Process.
Why it is a tell: Numbering reads as editorial rigour, so models number everything. Nobody navigates a landing page by section number.
The question: Does the visitor use these numbers to find or cite a section?

Tile pagination

Marker: Counters on images or bento tiles, like 01 / 4, when all four are visible at once.
Why it is a tell: It copies the chrome of a slideshow onto content that does not page.
The question: Can the visitor move between these items, and does the counter change when they do?

Generic step labels

Marker: Stage 1 / Stage 2 / Stage 3, Step 01, Phase II, Pass One used as the title of each step.
Why it is a tell: The model knows a process needs steps before it knows what the steps are, so the label is written first and the content fitted in after.
The question: What does the visitor actually do at each step, and is the number doing any work the order does not already do?

Word strip at the bottom of the hero

Marker: Three or four capitalised words in a row under the hero: TYPE / FORM / MOTION, DESIGN · BUILD · SHIP.
Why it is a tell: A portfolio cliché models reproduce to fill the space between the call to action and the fold. It lists disciplines nobody asked about.
The question: Do these words link anywhere or tell the visitor something the headline does not?

Scroll cue

Marker: Scroll, Scroll to explore, a bouncing arrow or a mouse-wheel animation at the bottom of the first screen.
Why it is a tell: It fills the lower edge of the hero with a gesture everyone already knows.
The question: Is there evidence that visitors of this page stop at the first screen?

Decorative status dots

Marker: A small coloured dot before nav items, list rows, badges or Available for work, often pulsing.
Why it is a tell: A status light suggests a live system. Models add it for motion and colour, with no state behind it.
The question: Which real state does this dot report, and what turns it off?

Pills laid on images

Marker: Tags or labels positioned over photographs: Brand · 02, New, Case study.
Why it is a tell: It layers the look of a tagged archive onto images that carry no tags.
The question: Does this label hold information the visitor cannot get from the image or its caption?

Middle-dot runs

seen in our runs Design studio landing: Design studio · three people · food & drink over the hero, found by audit.py.

Marker: Three or more items joined by middle dots on one line: Brand · Motion · Spatial · Lisbon.
Why it is a tell: The middle dot is the model's default separator for anything that looks like metadata, so it spreads to lines that are not metadata.
The question: Are these items a real list the visitor scans, and does the thesis ask for them on one line?

Decorative grids and crosshairs

Marker: Hairline grids, corner crosshairs or registration marks drawn behind content that does not align to them.
Why it is a tell: Printer's marks signal craft. Drawn over a layout that ignores them, they are wallpaper.
The question: Does any content sit on this grid, and does the thesis name the print reference?

Vertical text

Marker: A label rotated ninety degrees along an edge: INDEX OF WORK, SCROLL, a year.
Why it is a tell: Rotated type is the fastest way to look experimental, and models reach for it whenever a brief says creative.
The question: What does the rotation let the visitor read that horizontal text would not?

Micro-sentences under eyebrows

Marker: A short explanatory sentence between the eyebrow and the heading, or right under the heading, restating the section: "Everything we ship, nothing we promise."
Why it is a tell: Models pad each section with a line of framing because every section in their training looks framed.
The question: If this sentence were removed, would the visitor lose a fact?

Glow and neon

Marker: Coloured outer glows on buttons, text or borders: box-shadow: 0 0 40px ..., shadow-[0_0_40px_...], a neon text-shadow.
Why it is a tell: Glow reads as energy or technology without saying anything, so models add it whenever a brief sounds futuristic, playful or AI.
The question: What state or emphasis does this glow mark that the layout does not already mark?

Blurred blobs

Marker: Large blurred colour shapes behind the content: blur-3xl or filter: blur(80px) on an absolutely positioned element, or the same disc drawn with no blur: a positioned radial-gradient fading from a colour to transparent.
Why it is a tell: They fill an empty background with colour, so the page looks warm without a single decision being made.
The question: What would this section lose without the blob, and does the thesis name it?

Perpetual motion

Marker: Animations that never stop: animation: ... infinite, animate-pulse, animate-ping, repeat: Infinity, on elements that report no live state.
Why it is a tell: A loop makes a static page look alive, so models add pulsing dots and floating shapes to signal activity.
The question: What real, ongoing state does this loop report, and would a still frame say the same?

Reflex convergence

The test Did this choice (layout, type, palette) come from the thesis, or from the model's habit?

Automatic checks in this family: tell-equal-cards, tell-gradient-text, tell-glass, tell-em-dash. What they do not cover is a manual read.

Three equal cards

Marker: A three-column row of cards with the same size, the same structure and the same class list, usually an icon, a title and two lines.
Why it is a tell: It is the model's first layout for "features". Three is the number that looks complete, whatever the content.
The question: Does the content really come in three items of equal weight?

Repeated zigzag

Marker: Image left, text right, then the reverse, more than twice in a row.
Why it is a tell: Alternation is the cheapest way to make repetition look designed.
The question: Does each section's content call for an image beside it, or is the image there to keep the rhythm?

One layout family everywhere

Marker: Every section built from the same arrangement: centred heading, then a grid; or heading left, content right; section after section.
Why it is a tell: Once a model finds a structure that works it reuses it, since every section is generated with the same prompt pressure.
The question: Which layout family does the thesis name, and where does a section's content differ enough to need another?

Bento with an empty cell

Marker: A bento grid with a cell that holds only a gradient, a pattern or nothing at all.
Why it is a tell: The grid shape was chosen before the content was counted, so a cell is left over.
The question: How many items does the project have for this grid, and does the shape match that count?

Headline with a paragraph floating in the corner

Marker: A large heading on one side and a small paragraph pinned to the opposite top corner, aligned to nothing.
Why it is a tell: It imitates a studio portfolio header. The paragraph drifts because nothing in the layout asked for it.
The question: What does the paragraph align to, and does it need to sit apart from the heading it explains?

Hairlines on every row

Marker: border-t and border-b on each row of a list or specification table, so that every line is boxed.
Why it is a tell: Rules on every row are the default way to make a list look precise.
The question: Does the visitor need to track across columns here, and does the thesis name this ruling?

Score bars on a full track

Marker: Comparison bars drawn as a partial fill over a full grey track, for skills, features or competitors.
Why it is a tell: It lifts a dashboard component onto a marketing page to make claims look measured.
The question: What was measured to produce these lengths, and against what maximum?

Display face by reflex

seen in our runs Design studio landing: a book serif headline with two italic accent words.

Marker: A display face the thesis never names: an editorial serif (Fraunces or Instrument Serif most often), frequently with one italic accent word, or a heavy grotesque sans set tight and large.
Why it is a tell: Models map mood words to a face. Creative, premium or editorial calls up the free display serifs; clean, modern or SaaS calls up the heavy grotesque. Either way the face was summoned by adjectives, not decided, and removing one reflex tends to install the other.
The question: Which sentence of the thesis names this typeface, and says why this one for this product?

Warm paper ground

Marker: A faintly warm off-white background with near-black ink and a single red or rust accent, on a brief that never mentions paper, print or ink.
Why it is a tell: It is the current default for "tasteful". It looks considered on every brief, so it says nothing about this one.
The question: What in the product or the brief makes paper the right ground, and does the thesis say so?

Glassmorphism

Marker: Translucent panels blurring what sits behind them: backdrop-blur, backdrop-filter: blur(...) on cards, navigation or modals.
Why it is a tell: It is the quickest way to make a panel look current, so it lands on cards and navigation whatever the product.
The question: What does seeing through this panel show the user, and does the thesis name it?

Monospace everywhere

seen in our runs Design studio landing: mono micro-caps on the eyebrow, the navigation, the section names and the captions.

Marker: A monospace face on labels, navigation, captions and body numbers across the page, far beyond code and tabular figures.
Why it is a tell: Mono reads as technical and precise, so models spread it to look engineered.
The question: Which content on this page is code or tabular data, and does the mono stop there?

Gradient text

Marker: Headings filled with a gradient: bg-clip-text text-transparent, or background-clip: text in CSS.
Why it is a tell: A gradient fill is the quickest way to make a heading look branded, so it lands on headings whatever the palette.
The question: Is this gradient part of the palette the thesis validated, and does it survive the contrast check on every stop?

U+2014 (em dash)

seen in our runs Design studio landing: in the label over each project, found by audit.py.

Marker: The character U+2014 (em dash) in visible copy, typed or written as —, — or —.
Why it is a tell: Models join clauses with it far more often than the people whose writing they learned from. At that density it is a signature, not punctuation.
The question: Does the project's own copy, or the thesis, use this character? If not, what put it on this page?

Hollow copy

The test Does this sentence say something true of this project and of no other?

Automatic checks in this family: tell-filler-verb, tell-duplicate-cta. What they do not cover is a manual read.

Filler verbs

Marker: Elevate, Seamless, Unleash, Next-Gen, Supercharge, Empower, Revolutionize in headings and calls to action.
Why it is a tell: These verbs fit every product, which is why the model reaches for them when it knows nothing about this one.
The question: What does the product do, in the words its users would use?

One intent, several labels

seen in our runs Invoicing SaaS landing: Sign up free, Create a free account and Start free, upgrade later, all for the one sign-up.

Marker: Buttons and links that ask for the same thing under different words: Get started, Start free trial and Sign up on one page.
Why it is a tell: Each section is written on its own, and each gets a fresh phrasing of the same call to action.
The question: What is the one action this page asks for, and what does the product call it?

Poetic section labels

Marker: From the field, Field notes, On the bench, Loose plates over testimonials, articles or updates.
Why it is a tell: The model dresses functional sections in craft vocabulary to sound handmade.
The question: Would a visitor looking for this content search for these words?

Marker: Quietly trusted by, In good company, Loved by the best above a logo wall.
Why it is a tell: False modesty is the model's idea of confidence. The adverb carries the claim so that the logos do not have to.
The question: Are these logos real clients who agreed to appear, and what do they use the product for?

Register drift

Marker: A page whose tone changes between sections: technical in one, lyrical in the next, salesy in the footer.
Why it is a tell: Sections are generated one after another, and each drifts toward whatever voice its subject most often has in the training data.
The question: Which voice does the thesis set, and does every section read in it?

Unread strings

Marker: Any visible string (heading, label, alt text, button, footer line) that nobody read again after it was written.
Why it is a tell: Copy produced in one pass keeps the defaults of every family above. They surface on a second read and hide on the first.
The question: Has every visible string on this page been read once more, against the thesis, before calling it finished?

Sources: observations of landing pages generated by current models with no design guidance, and the tells catalogue of taste-skill (Leonxlnx, MIT), rewritten here as evidence and questions.