A universe of sentences, and a way to walk it
Behind the input box stand 589,444,128 ways of writing one thought — 476 voices, each scored across 52 dimensions of craft, crossed with tone, shape, size, vocabulary and approach. Choose one and a line comes back that has never existed before, in a voice you can name blind. This is the account of a place where original language is made at a scale no single writer has ever had time to reach.
What a writer is handed
None of this is inventory. It is the instrument itself — what a writer chooses from, and what the server has already settled on their behalf before a single word is asked for.
It knows what a model does wrong, and stops it
Everyone else sends the settings and hopes. Here the model is treated as a material with a known grain: it slides into its own habits, it reaches for the nearest word, it repeats itself the moment the framing holds still, it drops famous names, it answers a clinical seed with a dictionary entry. Every one of those has a countermeasure in the server, written on the day it was caught, and each carries the date it was caught on.
- It bans its own habits. Every 200 generations the server reads the last 2,000 lines it wrote, reduces them to word stems and bans its own top fifty — while exempting any stem you typed, so your own word is never taken away from you.
- That habit list is built from one kind of work only. Wish vocabulary was polluting it and making the ban depend on what sort of week you had had, so the sample was narrowed.
- The instruction's own words were found to be the contamination. Saying "aphorism", "wisdom" or "insight" inside the prompt pulls output toward greeting-card writing even when the seed has nothing to do with any of them, so the rules were rewritten to stop saying them.
- Three ban lists run at once — fifty words banned everywhere, ten to twelve tuned for each of the top two hundred voices, and a static list of what a model reaches for when it is not really trying.
- Six separate prompt families — standard, surprise, transform, refine, koan, fill-blanks. Not one template with switches. Each has structural requirements the others would break.
- Conflicts are settled before the prompt is built. About sixty parameters go through a precedence pass, so the model is never left to arbitrate between two instructions that disagree.
- Prompt strategies are numbered and tested against each other. Seven of them, with judge results from 317 samples showing two voices resist the technique that helps the rest — so the server routes per voice to whichever strategy that voice responds to.
- Every generation carries a nine-character code holding its full settings. In the last thousand rows, 824 have one. Paste a code back and the studio rebuilds that exact configuration.
- The code returns the settings but never the framing behind them. Reproducible for you; not extractable by whoever you send it to.
- The voice list was built by analysis, not opinion. 465 voices reverse-read into 9,300 craft traits, clustered into 52 dimensions, then 476 voices scored on all of them.
- Poor-fit options are dimmed, never disabled. Thirty-five percent opacity and pushed to the end, still one click away — because the unusual combination is often the interesting one.
- The filtering runs both ways. Choose a tone and the voices that suit it light up, not only the reverse.
- Even the word "aphorism" can be dropped from the brief, because asking for one biases the answer toward sounding wise. The commission itself is swappable.
- The vocabulary levels know their own rank. The three simplest override tone and sentence shape outright; the four richest serve the tone instead. The tone is in charge; vocabulary is the toolkit.
- The simplest level was moved from age seven to age nine because age seven turned every object into a toy or a ball, and a spinner stopped being a spinner.
- Asking for four versions runs a real divergence pass — one call opens the seed into ten angles, a selection step keeps the four furthest apart, then four calls run with length, register, word tier and temperature each fixed differently so they cannot drift back together.
- Losing one of four returns three. Written into the code with its reason: you waited eighteen seconds, and losing one is recoverable while losing all of it is not.
- Asking again for the same thing makes the engine bolder — the temperature climbs with each repeat instead of serving the same answer twice.
- The previous answer is handed back as something to avoid when you re-send the same input.
- A near miss is repaired, not rejected. Output that almost meets the constraints gets one corrective call, recovering about half of what would otherwise be thrown away.
The rules are checked after the answer comes back
A rule inside a prompt is a request, not a guarantee. So the parts that carry legal or human risk are checked twice — once in the instruction, once on the way out.
- A detector reads the finished line for real people's names, because the rule that forbids them is not trusted on its own.
- A refusal never renders as writing. The exact phrases a model produces when it declines are recognised and handled.
- Developer notes are stripped before send. The comments that explain the prompt to a person never reach the model.
- Outbound text is scrubbed for anything secret, kept as a second defence even after the key was moved off the client.
- Every response carries provenance marks at the start, at the end, and inside random records of any list.
- Sacred input is identified by tradition, not by the longest word matched. "Vishnu lessons from Buddhism" resolves correctly because the traditions are scored against each other.
- Each sacred entry names the voices allowed near it. The satirist, the cynic and the comedian cannot reach it at all.
- Honorifics are filled in where they were never written, and forbidden framings are inherited by whole categories rather than typed entry by entry.
- Greeting-card phrasing is caught separately, because it passes every structural test and still reads as a card bought at a petrol station.
- Low confidence no longer means silence. The tool used to return three fixed refusal cards without calling the model at all. Now it writes, and marks the result for checking — giving up before trying is not respect.
- Each tradition carries examples set at its highest bar, so the model has something to reach for instead of an average to settle into.
Settings no other writing tool offers at all
Anyone can offer a tone menu. These controls come from a different question altogether — not how should this sound, but who is receiving it, from what position in time, and out of which field of life.
- The reader's age, at both ends of life. Under seven, eight to ten, eleven to fourteen — and sixty to seventy-five, and seventy-six and beyond. Most tools build for the middle and stop.
- Direction and time as separate settings. Think, feel, warn, remember, laugh, challenge, comfort — crossed with now, before, after, always. Neither is a tone.
- A field of life as the source of imagery. Pick medicine and the line borrows from medicine; pick farming and it borrows from farming, for that one generation.
- A word can be required, or required to stay out. In the second setting the word is the seed of the thought and must not appear in the sentence.
- You can add a voice. One command builds the profile and the new voice takes its place beside the 476 already there.
- The input box is a command line. Random word, harder random word, a phrase, a subject list, a photo brief, wildcards, blanks to fill, and square brackets that override the whole left pane for one line.
- Random never repeats itself twice running on any setting — randomisation with a memory.
- Clashing settings are named and offered a fix rather than quietly producing a compromised line.
- When the server reduces your sentence count because the size is too small, it says so, instead of leaving you to think you were ignored.
One word, four briefs, four different lines
Here is the whole argument in four lines. The seed is one word — silence — and it does not change. Only who the line is for changes.
| What it is for | Who reads it | What came back |
|---|---|---|
| nothing given | nobody given | When everyone stops talking, your next move appears. |
| headphone brand | premium buyers | Wear less noise. Let every detail speak for you. |
| meditation app | stressed millennials | Let the noise pass. Notice the space, return. |
| a condolence card | someone who lost a parent | When silence visits, we honor what you lost. |
The fourth row is the one to look at. Nothing in the request said stop selling. The engine understood that a condolence card is not a product and dropped the sell entirely. Three more briefs, same day, same engine: a perfume gave First rain on warm skin, gift of instant relief; a page-not-found screen gave This page vanished, so your next click must find it; a thirtieth anniversary toast gave Thirty years side by side, love still learning.
The same thought, in voices you can tell apart blind
One seed — a tool that remembers what you rejected — through four voices. Nothing else changed.
| Voice | What came back |
|---|---|
| George Carlin | Keep rejected notes; that tool remembers, and it bills you later. |
| Basho | Rejects logged by the tool. |
| Rabindranath Tagore | A rejected nail hums in the tool, and later, it holds. |
| Richard Feynman | Your drafts live on, but your rejections return as measurement. |
Carlin bills you. Basho takes five words. Tagore finds the nail. Feynman turns it into measurement. That separation is what fifty thousand corrections buys, and it is the hardest thing on this page to fake.
What happens after the first line
A generator hands you a line and leaves. Here the line is the beginning: two thoughts can have a third, a sentence can be traced back through three civilisations to the people who felt it first, and the tool knows when you have adjusted enough.
- Two lines can produce a third. The specification writes down the failure beside the success: joining the two sentences is wrong, and the result has to read as though it were written on its own.
- The waiting is deliberate. Seven drawn figures think on screen and the answer is held until the drawing finishes — fifteen seconds the first time, three by the third round, because the work gets faster as the conversation warms.
- Adjusting is capped at three rounds, with the counter visible. After that you accept it, edit it yourself, or start again.
- Precedent is looked for in three regions — South Asia, Asia and Africa, and the West — and that split cannot be changed, because it is the statement.
- You can choose without typing. Seven rounds of two lines, five seconds each, and your picks become the seed.
It tells you what it actually cost
A tool that spends your money on your behalf owes you the truth about it, in the moment, without being asked. These five are that debt paid.
- Three quality levels, each with its real price on the label. Sketch ten lines for a penny, and step up only for the one you are keeping.
- Every displayed price carries a forty percent margin, held on a single line of code, so the number is a ceiling rather than a flattering floor.
- When the cheap model is busy and the request falls through to an expensive one, the cost line says so. It names what actually paid.
- Deleted work is kept with its reason. A rejected line stays on file with every setting that produced it — the material for finding out what "not good" means. Recipients' names in wishes stay in your own browser and never reach the server.
- The record is the server, not the tab. 95,490 rows, searchable, exportable behind a password, with a settings library built from your own most-used combinations.
Half a billion settings, before a word is typed
476 voices, 54 tones, 26 sentence shapes, 21 kinds of creative approach, 7 vocabulary levels, 6 sizes. Multiplied, that is 589,444,128 distinct configurations — and that is before a seed word, a reader age, a direction, a moment, a field of life or an anchor is chosen, and before the framing itself varies. This is not a menu. It is a space.
A great writer leaves a few thousand lines worth keeping, and it takes a life to do it, because a life is the rate at which a person can think. This thinks at a different rate. The keepers are found by reading rather than by waiting, and you may set the bar wherever your judgement puts it — one in fifty, one in two hundred, one in a thousand — because the arithmetic arrives at the same place from any of them: more original, keepable short-form work than any single writer has left behind, from a machine that has not finished its first year.
And it is original by construction, not by luck. Cliché is banned by list. The tool's own habits are banned by measurement. Famous phrasings are pushed away by the rule that forbids the memorised quote, and real names are refused outright. What comes out cannot be a remembered line, because every mechanism in the server is arranged to prevent exactly that.
Why this is hard to copy
Any team can put a tone menu in front of a model. What cannot be shortcut is the list of defects — the drift, the obvious word, the sameness under a fixed framing, the reach for a famous name, the dictionary answer to a clinical seed — because that list is written by being wrong in public, repeatedly, and writing down what happened each time. More than fifty-five thousand inputs across a year is the price of it, and there is no shorter route. What sits in these three tools is not prompt engineering. It is a materials record for a material nobody else has bothered to measure.
And these three are not the practice. They are one instance of it. The same method built more than fifty other tools on this property, each one arriving at behaviour the companies making the models have not thought to build — which is what happens when the person writing the corrections is the person who has to live with the output.
Original pedagogy, voice, and design — all rights reserved.