Goals

The toolbox is for every assistant now, not one. AGENTS.md leads, because it is the open standard the Linux Foundation stewards and Codex, Cursor, Copilot, Gemini CLI, Aider, Windsurf and Zed all read it natively. Then the tool-specific ones for anybody standardised on something: a Claude Code skill that becomes a command, a CLAUDE.md paragraph, a Cursor rule scoped by glob, a Copilot instructions file, and a plain brief for anything that reads English. Six forms of the same thing, each whole, each with a button that takes all of it.

It was three stacked bands of colour, which the Commander called a pool towel, and he was right. The fault was not the colours: stacking equal horizontal strips is what you do when you have not decided what the thing is. So the page decides. French Press came out of a cupping form and already draws in line, and a program is a path something travels down, so the tool is a figure now — the request entering from the left, five numbered gates that can each stop it, the host at the other end. The numbers in the drawing are the numbers of the rules beneath it, generated from one list, so the figure cannot drift from the program. It already caught itself: the list and the drawing had fallen out of order and the build showed it immediately.

The toolbox is built for how people actually work now rather than as a page about a program. Almost anybody who wants this is in a terminal with an assistant beside them, so the page leads with the three things they can take: a skill file that drops into a repository and becomes a command and travels with the repo, a line for the standing conventions file so it is true for every task instead of remembered for one, and a plain brief for any assistant anywhere. Each one is whole, each has a button that copies all of it, and the source sits underneath for anybody who would rather read it than run it.

Code got its own face. Everything on this site was set in one of two voices, and something a reader is going to paste has to look like code and keep its columns, so there is a third now — a system stack, costing no request. The measure gates leave code alone for the same reason: a line rewrapped to a comfortable seventy-five characters is a line that will not work when it is pasted, which is the only thing anybody wants to do with it.

We were rate-limited, and that is not a scheduling problem, it is a diagnosis: a person reading a forum is never told they are asking too often, so being told it means the thing asking was not behaving like one. The manners now live in one small compiled program that everything here has to go through. One request at a time across every tool, twenty to forty seconds between requests to a host, sixty a day and then that host is finished, all of it remembered on disk between runs, robots obeyed, and a refusal that puts a host away for six hours with no retry and no second agent. A check refuses to build the site if any tool reaches the network on its own again.

The first version of that program did not work. Its robots reader took a rule's trailing comment as part of the rule, so a line reading Disallow slash works question mark, hash, cruel but efficient never matched anything — the one thing it exists to do, it was not doing. Both versions are on the toolbox page with the diff between them, because that is the part somebody else would actually need.

The wall glitched while scrolling. It was a column flow, and a column flow reworks the whole layout every time a lazily loaded picture arrives, so the page jumped under you the whole way down. It is a grid now and each cover reserves its exact shape before it loads, passed in as data rather than guessed. Every form still keeps its own proportions and nothing moves.

How we read a room is written down on the About page, because it is the part that has to be right rather than the part that is clever. We read what is public, one page at a time, at the speed a person reads. We never join, post, reply, vote or ask. We never name anybody and we never quote a comment — a quotation is a person's own words and a person can be found by their words even with the name taken off — so what goes up is what the room worked out, in our sentences, with a link to the room so you can check we read it right. Nothing from a room ever moves a score.

The search now knows which rooms suit which kind, because a novel is argued about somewhere different from a film and a manga somewhere different again. It found nothing today: Reddit answered too many requests and we stopped, which is what that answer means. The reading resumes when it is welcome. Writing that here rather than quietly working around it is the point of having written the method down at all.

The visual ones are in, and they use what the site already owns rather than borrowing somebody else's gag. Type argue and the shelf comes apart by how far each film's critics were apart, so a film they all liked the same amount barely moves and a fight swings out of line. Type lights out and the page goes into the instrument room — nothing was built for that, it is the second livery this kit has always carried. Type quiet and every colour goes, and the four kinds of argument become impossible to tell apart, which is this site's own rule about colour demonstrated on the reader instead of explained to them. Eighteen hidden answers besides, every one worked out from the shelf as it stands so a joke cannot outlive its fact.

The Overthinkers. Not theories: a theory predicts something and can be killed by evidence, and almost nothing in that genre can. A case is a question about a work that can be argued out of the work itself, laid out the same way every time — why it is worth asking, what holds it up, what pulls it down, and where it stops. The two columns are the same width and always will be, because anybody can write the case for something and printing the against beside it is the whole difference between thinking and advocacy. Two cases up: whether Orbital is a novel at all, and whether the critics who disliked Thunderbolts disliked the film or the thing it is attached to.

The wall stops pretending everything is a film. Every cover was being cropped to a one-sheet's proportions, so a book jacket, a manga volume and a series title card all came out looking like films, which is the opposite of what a shelf holding four forms should do. They keep their own shapes now, and what a thing is is printed on its tile — but only when it is not a film, because a FILM label on a film is a word that changes nothing.

There are things hidden in the search field. Every one of them is worked out from the shelf as it actually stands, so a joke can never outlive the fact it was made about, and every one is about what this site is rather than a gag stuck on top of it. There is one on the keyboard too, and what it does is take the number off every tile and show you the argument underneath it instead.

Marlow has a shelf. Seven things he keeps, with a line each about why, and no number anywhere on the page — nothing on it is scored, nothing on it is read for the site, and none of it goes into any average. It is not in the navigation on purpose: a shelf is something you find in somebody's room, and the way in is from his name on a reading.

French Press is not only films. Four works in two new kinds: Orbital, Intermezzo and James, and Suzume. The method did not change because it never had anything to do with film — a novel gets reviewed in sentences too. The harvester had to learn one thing: a film piece says David Rooney of The Hollywood Reporter wrote, and a book piece says Writing for The New York Times, Joshua Ferris said. Same evidence, other word order. Forty-two works, 150 reviews.

One question is now asked before any source is chosen: if a thousand angry people arrived at this link tomorrow, who would be standing there? A book, a studio, a newspaper, a board, a critic paid to be read, or a room with no door on it can take that. A fan writing under a pseudonym cannot, never asked to be read here, and could not make it stop. Never a fan author, never a single fic, never a wiki editor's name, never a single comment, never a profile. It is written down in data/exposure.json, it is on the About page in plain words, and a program checks the built site against seventeen shapes of link that mean one person rather than a room. Nothing publishes if it finds one.

The writing was a metronome. Two sentences, break, two sentences, break, all the way down every page, so a passing remark got the room of an argument and the argument got the room of a remark. It reads as a voice, which is why it survived every pass I made over it. It is sameness. The About page is rewritten with the length carrying the weight of the thought, and there is a measure now: the spread of paragraph lengths and of sentence lengths across a page, wired into the gate. Writing that reads flat does not get published.

The type scale was the reason pages only looked right when you zoomed. Body text sat at seventeen pixels whether the screen was 390 wide or 1440, while the display sizes climbed to sixty-five, and the About paragraphs ran to eighty-five characters a line. Zooming fixed it because zooming is the reader doing by hand what the scale should do: raising the size and narrowing the measure at the same time. Body grows with the viewport now, the hero comes down to meet it, and a line that runs past eighty characters on a wide screen fails the build.

We stopped claiming we can spot a bought review. The site said it removed reviews that had been bought or arrived in a pile-on, and that was on the front page, in the description every search engine reads, and in Marlow's own account of the method. It is not something anybody can do: a paid review written by somebody competent reads exactly like an honest one. A review is not scored for one reason now and it is a reason about us — our source handed us too little of it to place.

And we stopped drawing those critics as rejects. A review we could not score was pushed to one side of the page, greyed, given the colour that means rejected and struck through with a line drawn across the writer's own sentence; one critic's page said he was the one critic whose review we had refused. He did nothing — an encyclopedia quoted five words of his review and five words is what reached us. The limit was ours and we drew it as his.

Any site can carry an FP score now. Every film page shows the card working and the one line that puts it in your page. No script in it, nothing to sign, nothing to pay, and it always links back to the reading, because a number nobody can check is a rumour.

Two things on a film page were saying something untrue. The guide printed the board's word beside the row it named — Violence, then \u201cviolence\u201d — which is a tautology, and it pulled critic sentences in by keyword, which is how \u201cmade a bleak subgenre of cinema a little bit sunnier\u201d ended up filed under Upsetting. A row is a name and a bar now. The sentences are all still on the page, in full and attributed, where they were always going to be read properly.

\u201cWorst thing said\u201d was a warm review on fourteen of thirty-eight films. Hamnet's was an 86. Spider-Man's was \u201cdark and thrilling new chapter\u201d, two points under a review calling it fantastic. Nobody said a harsh thing about that film, and a heading claiming somebody did is a lie told for a symmetrical layout. The pair only appears when the critics were more than ten points apart, and it is the kindest and the harshest, which is true at any level. Otherwise there is one row and it says what they said.

The guide is a bar. It was three sentences of boilerplate repeated seven times down every film page, and it read as a wall of the same words. The bar is not ours: it is the classification board's own five steps, in its own order, very mild to very strong, so the drawing and the ruling cannot disagree. Green to red, five cells with the dividers showing, and the board's word printed beside it so the level never rests on colour. One line at every width, because a strength bar that wraps has stopped being a bar.

Marvel is on the shelf: Thunderbolts, Captain America Brave New World, The Fantastic Four First Steps, Spider-Man Brand New Day. Thunderbolts is the interesting one. Two critics, 57 points apart: one says the cast finally works together in a way the studio has not managed in years, the other says it is dollop upon dollop of dourness and that the end-credits tease is the sales pitch. Thirty-eight films, 136 reviews.

A phone is not a narrow desktop. On a film page the answer comes first now — the score, how far apart the critics were, and the title — and the poster follows at a size that supports it instead of being a wall you scroll past to find the name of the film. The critics roll becomes a list rather than a column of cards: 101 cards was nineteen thousand pixels of scrolling. Anything you tap is at least a finger tall. The tablet gets its own counts too.

A five-word sentence spelled down five lines, because the column it was in is 168 pixels wide. Every page is measured for it now at three widths. The device check itself was broken twice while finding this: it wrote its probe copy to the wrong directory, so it was measuring pages with no stylesheet, and it could not hold a viewport width, so it answered for 500 pixels when asked for 390. Both fixed; the page under test now renders in a frame of exactly the width being asked about.

Every page is measured on a phone, a tablet and a desktop now, and the measurement is a gate. kit/devices.py loads every page at three widths and checks whether the document is wider than the screen it is on, then names the element sticking out. It immediately found one: the front page ran eight pixels wide at every size, because a full-width band inside the page column reached back out using 100vw, and 100vw counts the scrollbar. Invisible in any screenshot; a page that rocks sideways under a thumb on a real device. The band is a sibling of the page column now and the arithmetic never comes up.

Guides: 176 of 238 rows say something specific, and 29 of 34 films have a board record or critic detail. For the five with neither, the page says so plainly and names the boards we checked. A film that has played festivals and had a limited run often has no certificate anywhere yet, and 'no board has rated this' must never be allowed to look like 'there is nothing here'.

The Compare page was a wide table on a phone, running off the right edge with the band and the review count cut off. Under 680 pixels a row is a card now: poster, score, title, and how far apart they were, stacked. Nothing is dropped and nothing moves sideways under your thumb. The spec pairs on a film page stack again too; they had stopped when the rows lost their rules.

The guide has something in it. It was 22 filled rows out of 238, because the only text we held per film was a critic sentence or two, and a critic writing about Hamnet is not trying to tell you whether there is swearing in it. The body whose entire job is that question publishes the answer: the British Board of Film Classification gives, for every film it certifies, the sentence that earned the certificate, in the same vocabulary every time. That is the spine now, printed in the board's words with the board named and linked, and the critics are the second layer. 111 rows of 238, and 22 of 34 films have a board record.

We do not print a certificate. It lives in an image on the board's page and both ways of reading it handed back a U for a film rated 15. A wrong certificate is worse than none, so the page says plainly that we could not read it.

The front page arrives in one screen. Four films across the top, one from each kind of critical argument, each on the colour for how much the critics agreed, and the shelf starts underneath them. It was four stacked full-width banners, which is the shape a shop front uses because it has four things to sell; this has thirty-four films and a reader who came to look at them, and the stacked version cost 1,280 pixels of scrolling to get past four of them.

A pill button had turned up in the front page, borrowed off a site we were studying for its composition. FP has three radii and each one means something: nothing at all, one pixel on a drawn bar, two on a control you type in or press. A fourth now fails the build. The action is a stamp instead: a solid block of the field's own light, squared off, which is what belongs on a form.

Every critic has their own page now. The roll is a mark, a name, a publication and the reach of what they gave; what we have to say about somebody lives on their page, where there is room for it. The mark is not a face and never will be: it is a paragraph reduced to the measure of its lines, with one line heavier because that is the line a review turns on, and its shape is fixed by the person's name. A hundred and one critics, a hundred and one different marks.

Oh so many hairlines. One page was painting a hundred and eighty-three grey rules, one under every row, because a rule is a way of separating things without deciding which of them belong together. They are gone: four remain on a page, and all four are chrome. Space does the grouping now.

The critics page is now a roll of the writers this site is built out of, and it exists to send readers to them. Nineteen critics we have read more than once get a line saying what we think is good about them, and every one of those lines can be argued with using the numbers on this site. The eighty-two we have read once do not get a line: one review is not enough to write a character reference, so their own sharpest sentence stands there instead. Every publication links out. Disagree with any of it: [email protected].

There is a feed at /feed.xml. Every reading as it lands, for any reader app that wants to carry these. Nothing behind a login, nothing costs anything, and the credit belongs to the critics rather than to us.

The dark room's ground is now derived rather than chosen. It was #141210, a warm near-black picked by eye with nothing behind it. It is now the roast's own hue, 21.7 degrees, held down to 6.5% lightness: the coffee with the light taken out. Twice the warmth of the old value, and every step of it can be checked.

The contrast tool held its own copy of the dark ground, so when the ground moved it went on measuring a colour the site had stopped shipping. It reads the ground from the room now. Eight ratios in the room were wrong and are corrected; every hue still clears 7 to 1.

Corrected: Deadline Hollywood's critic was on the site twice, as Pete Hammond on Hamnet and Peter Hammond on Jurassic World Rebirth. The encyclopedia's prose says Peter and its own citation for the same sentence says Pete. He is one person and now reads as one. His citation number followed the correction rather than a new one being issued.

Threads turned out to be readable without registering an app at all, which was the whole of the blocker on goal 02. Thirteen films now have a conversation found and waiting. None is credited yet: a person has to read the thread and write what it actually worked out, and that is the part a machine must not invent.

Twelve wide releases read and scored, bringing the shelf to 34 films and 128 reviews. Marty Supreme was harvested and did not make it: after one quote turned out to name two other films instead of this one and could not be placed, a single review was left, and one is not a reading.

Posters corrected. The picker was taking the first image on an article, which on a wide release is a photograph of the director; it now takes the poster the article itself puts in its infobox and refuses anything that is not portrait. Twenty-nine of thirty-four films, owner named on each.

The desk is now a stylebook: every ruling on how FP writes, numbered once, dated, and superseded rather than rewritten, with the word list the build enforces published under it.

About answers thirteen questions about the whole site, each one linkable, and is in the header on every page.

Marlow Vance takes the byline. Every film is signed and the About page is his.

Posters live for 17 of 22 films, fair use, owner named on each. One had no owner recorded, and its line says so.

The guide goes on every film: seven categories, a level each, and the critic's own sentence underneath.

Reddit reachable through its public feeds after the API refused us. 13 of 22 films have a thread worth reading.

Twelve more films read and scored by hand. 22 films, 82 reviews, 72 critics.

First three films read. The President's Cake lands at 67 against 99 percent elsewhere, because two substantial dissents count at full weight here.