Findings
Every UX failure and gap in the register, open first, then by severity. A failure works against the reader; a gap is something the reader needs that is not there. Each status says when it was last checked, and on which version: nothing here claims to be current beyond that.
Filter the register
- The first phone screen explains how the site is built, not what it is about
A contact who taps a link cannot tell in one screen that this is a publication about AI agents, security and publishing, or why they were sent it.
- The home page is 59 phone screens long, with 845 links
The front page tries to show everything, so it chooses nothing for the reader. Nobody scrolls 59 screens; the curation is invisible.
- All three rating scales start on 3
Neither the reader nor the author can tell "not rated" from "useful". Every rating collected so far is ambiguous.
- Ratings never reach the author
The first contacts sent to the site teach the author nothing about what landed.
- Opening an article charges a penny with no notice
The first thing a new reader learns about the reading meter is that their balance dropped without being told.
- Nothing shows real money moving
Readers and publishers asking "is this a real business?" find no answer.
- It is unclear how a writer gets paid
Writers and publishers, the people the meter is pitched to, cannot see where their share comes from.
- No contact for a pilot on the Replicate page
The one reader ready to act has nowhere to go. A blocking finding of the synthetic-user study.
- Two domains claim to be canonical for the same article
Search engines split the article's standing between two copies, and readers can land on either.
- The founder page says "the picks above" and no picks render on a phone
The page a founder is sent to points at content that is not there, on the first visit, which is the only one that matters.
- Setting location.hash inside a vault app re-navigates the frame: every link silently dead
An app that works when opened as a file ships to readers with no working navigation, and nothing errors.
- Scrolling back up a little puts focus in the address bar and opens the keyboard
The reader was not asking to go anywhere, and loses half the screen to a keyboard. Worse, it looks like the page has a bug: the first reading was that the site had somehow started an edit. The browser's fault is charged to the site, and the reader's concentration is broken either way.
- A security prompt that asks you to approve a command it does not explain, and cuts off
The person is asked to make a security decision and given only code to make it with, partly hidden, on a phone. In practice the choice is to stop the work or to trust blindly, and every prompt like this teaches the habit of pressing Allow without reading, which defeats the point of asking. Here the command was harmless (a one-line edit to two files), but the prompt would have looked the same if it were not.
- One unlabelled X throws away the whole recording, with no confirmation
A minute of dictation, often the hardest part of the message, is gone in one tap, on the button that looks least consequential. It has already happened to the reporter: "I literally fell for this in the past."
- The live transcript sometimes stops while you are still speaking, and nothing says whether it is still recording
The reader has to keep a mental check running while they speak, watching for words to appear, and stops trusting the tool whenever they do not. When it fails the recording cannot be recovered or resubmitted, so the loss is total.
- Content is hidden until JavaScript runs
On a slow phone, or for any reader or agent that does not run the script, the page is blank.
- Two navigation bars, and a 60-link menu that mixes reading with building
The reader has to sort the publication from the workshop themselves, twice.
- Missing pages show the default GitHub Pages 404
A mistyped or old link ends the visit on a page that does not even say which site it is.
- One topic holds more than half the articles
A CISO who picks the obvious topic still faces 35 long reads. The classification does not help anyone choose.
- Tags are not links, and there are no tag pages
The most precise way into the archive looks clickable and goes nowhere.
- Filters cannot be shared
You cannot send a contact "the security articles"; you can only send the whole list and an instruction.
- Series are invisible
Readers land in the middle of a story without knowing there is a beginning.
- There is no "start here" path
A newcomer facing 73 long articles has no way to know which five to read first.
- Cards do not show reading time or format
A reader with three minutes cannot tell a note from a 40-minute analysis before opening it.
- The best reader feature is buried under metadata and a long italic abstract
The short version that would let a reader decide is found only by those who already decided to read.
- The byline names one human; the AI authorship is in the footer
A reader who finds the disclosure late reads it as something hidden, on a publication whose pitch is provenance.
- There is no page that says why the publication exists
Readers cannot judge the argument without knowing who is making it and why.
- Articles join the reader's list without being asked
The list stops meaning "what I chose to keep" and becomes a history the reader did not ask for.
- "Credit can go below zero" reads as a debt
Readers hear "you can owe us money", on a site that has not asked them for any.
- Three different article counts
When a site cannot count itself, the reader doubts every other number on it.
- Opening a persona link adds it to the browser without asking
The reader did not choose to become "a founder" on this site; it was done to them.
- Persona pages are a list, with no story, opinion or next step
A persona link is a promise that the site was made for this reader; a list does not keep it.
- localStorage throws inside the vault app frame, so reader data is lost on reload
A reader's notes disappear on reload, the worst failure a note-taking feature can have.
- The toolbars come back on any small scroll up, taking screen and moving the page
The reader scrolls up a line to re-read a sentence and the whole page jumps, the text moves away from their eyes, and a toolbar they did not ask for takes a ninth of the screen. Each time, attention goes to the chrome instead of the text.
- Whether the toolbars hide depends on how fast you scroll, not how far
The reader cannot ask for the state they want. To read with the toolbars hidden they have to flick, lose their place, find it again carefully, and hope the correction does not undo the result. Every attempt takes attention away from the text, which is the opposite of what hiding the toolbars is for.
- Three voice controls with no labels: the safe one and the send-as-is one sit side by side
Every first use is a guess, and the guess has consequences: a wrong tap either destroys the recording (UX-060) or sends a message full of transcription errors to an agent that will act on it.
- The version number is shown three times on the first screen
Engineering detail repeated where a reader looks first; it reads as a demo, not a publication.
- A "For agents /llms.txt" strip sits above the masthead
Useful to agents, noise to a person, and it takes the most valuable line on the screen.
- The editor's note is written for someone who read the last two editions
The one personal voice on the front page talks past the newcomer the page most needs to win.
- Links without extensions in one article
They depend on the host resolving extensionless paths; on a strict host or a mirror they break.
- Each article ends with seven next steps instead of one
At the moment the reader is most ready to continue, the page makes them choose among seven.
- Search only matches titles, teasers and tags
About 300,000 words are unsearchable; a reader who remembers a phrase cannot find the piece.
- Author tooling is shown to every reader
A tool for the author, printed on every reader's copy.
- There is no way to say what was missing
Scores say what, not why. Requests that would become the next pitches have no channel.
- "You pay the price" reads as a warning
The feature (your rating sets what the page costs you) is a benefit; the wording makes it a threat.
- The feed is titled for the old site
Subscribers see the wrong publication name in their reader.
- The study page reports a fixed bug as current
A visitor reading the study believes the site is broken in a way it no longer is.
- A vault app can load images and scripts from a site but cannot fetch its data
Builders reasonably expect CORS-open data to be readable and discover otherwise in production; apps must snapshot data instead.
- Query strings break inside the vault host, so filters must live in the route path
The usual way to make a filter linkable is not available; each app has to invent a path encoding.
- A highlight across paragraphs is cut at the first paragraph
The reader loses part of what they chose to keep, without being told.
- Search results show no snippets
A reader cannot see why a result matched, so cannot judge it before opening.
- Stacked tables on a phone lose their column labels
Values without labels on a phone: the reader has to remember what each line means.
- "Ask the article" has never run against a real on-device model
The feature is shown to readers on a path nobody has seen work.
- The transcript is cleaned up in a second pass without saying when, or what changed
The reader can no longer tell their own words from the model's interpretation of them, so they cannot tell what to check. For an instruction to an agent, a quietly changed word can change the task.
- The £5.00 balance chip has no label
An amount of money with no explanation reads as a bill or a trap; the synthetic-user study flagged it.
- The top-up page said "simulated" while asking for money
A payment page that says it is not real is a payment page nobody can trust.
- Articles were published on two domains, and diverged
The same publication was more complete on the wrong site.
- 51 vault images missing across 10 articles
Evidence images are the point of an evidence-first article; broken ones undo the claim.
- Persona links were designed but not live
Every contact got the same front page, made for nobody in particular.
- The masthead line describes a method, not a subject
The one line every visitor reads tells them how the site works instead of what it is about.
- The edition label is older than the articles under it
A front page that looks three days stale on the day it was published tells a reader the site is not kept.
- File-path pills scrolled the whole page sideways at 390 px
Sideways scroll on a phone makes every line harder to read and hides the edge of the content.
- Re-rendering the whole view on every save lost the reader's place
A reader who highlights a sentence is thrown somewhere else in an 8,000-word article.
No finding matches this filter. Show all.
For things that already get it right, see what good looks like. Browse the same register by product, problem, visibility, recommendation and principle.
The register is one file, /data/findings.json. This page, every finding page and every count on the site are generated from it. To add a finding, see the method.