We read 77 outside sources: research papers that measured AI writing, working writers who use these models, the tricks people use to get better prose out of them, how you can tell whether an essay is good, and classic advice on what makes an essay worth reading. This page sets that against our own two studies — the writing gap (what's wrong with Ink's prose) and the writing loop (sixteen attempts to fix it). It answers three questions: what did others already find that we found too (§3), what does the research say about the fixes we already tried (§4), and what haven't we tried yet (§5).
Short answer: the blandness comes from training. It's already there before our prompt is read.
After a model learns language, there's a second training step. People rate its answers, and the model learns to give the kind of answer people rate highly. The trouble is that raters tend to prefer what sounds familiar. So the model learns to reach for the most familiar way of saying things, every time.
| what was measured | what they found | who |
|---|---|---|
| The same model, before and after the rating step: how different are its answers from one another? | After the rating step, its answers were about half as varied. | Kirk et al., ICLR 2024 |
| Ask the same question 200 times: how alike are the 200 answers? | Before the rating step the answers overlapped 16%; after, 35% — twice as samey. | Mohammadi, 2024 not peer-reviewed, one model |
| Why does the rating step do this? | Raters prefer familiar-sounding text, and the maths shows any such preference pushes the model toward its single most likely answer. | Zhang et al., 2025 not peer-reviewed |
| People wrote essays with help from two AIs: one before the rating step, one after. Both wrote about the same amount of each essay. | Only the after version made the essays sound alike. The humans' own sentences didn't change — the sameness came from the machine's words. | Padmakumar & He, ICLR 2024 |
| Can you undo it at writing time? | Only partly. It's built into the model itself. | Karouzos et al., 2026 not peer-reviewed |
A word list says "the AI uses delve, tapestry, pivotal." A way of talking is the whole manner: how formal, how certain, how puffed-up. It's like the difference between telling a friend something and reading a press release. You can hear it even when no single word is wrong.
Kobak et al. (Science Advances, 2025) proved this with a counting trick borrowed from "excess deaths" in a pandemic. They tracked how often every word appeared in 15 million medical paper summaries, year by year, used the trend to predict 2024, and looked at which words showed up far more often than predicted. (Delves came out 28 times over.) Then they compared two sudden shifts:
A person would write "We looked at how it works." Swapping one word leaves the puffed-up manner in place. Our own study found the same thing: Ink's worst habit isn't a bad word, it's a manner — short, certain verdicts like "That is the whole game."
Each AI piece, read on its own, can look original. The problem shows up when you put many side by side.
Short answer: expert readers reject AI prose about 5 times out of 6. Ordinary readers often prefer it. That's why the expert's read is the one that counts.
The best measurement is a 14-question checklist written by professional writers — "does the ending feel earned?", "is there anything left unsaid for the reader to work out?", and so on. Ten expert judges applied it to 12 New Yorker stories and 36 AI stories written from the same plot summary, at the same length.
| share of checklist questions passed | New Yorker | GPT-4 | gap |
|---|---|---|---|
| all 14 questions Claude 30% · GPT-3.5 9% | 85% | 28% | −57 |
| Is there anything left unsaid for the reader to work out? | 89% | 11% | −78 |
| Does the emotion move and change, or stay on one note? | 92% | 19% | −72 |
| Does the ending work? | 92% | 19% | −72 |
| Is the shape of the piece original? | 64% | 8% | −56 |
| Do the people in it develop? | 61% | 17% | −44 |
| Is the thinking original? | 92% | 44% | −47 |
Chakrabarty et al., CHI 2024. The judges agreed with each other well on overall scores, less well question by question. Each story cost about $20 and half an hour of an expert's time.
A second study by the same lead author: 28 trained fiction writers against three AI models, judged blind in pairs. When the AI was simply asked to write in a style, expert judges chose the human 83 times out of 100.
Short answer: most of our 29 habits were already known under other names. Our top two weren't.
The best outside check is a 2025 study where 18 professional writers hand-edited 1,057 AI paragraphs — 8,035 edits — and sorted every edit into seven kinds of fault. Our four readers, working with no list, found the same faults.
| our habit | what others call it | who |
|---|---|---|
| "Not this, but that" | the most-mentioned AI habit among working writers. One writer bans it by name in the instructions he gives the AI. | Kriss (NYT Magazine), gwern, others. Nobody has actually counted it — widely noticed, never measured. |
| Nothing is wasted · the reader is put in a lesson | explaining what doesn't need explaining — 18% of the professional edits | 2025 editing study |
| Costume words · piling on intensifiers | overwritten prose and awkward word choice — 28% of edits, the biggest group | 2025 editing study; Kobak's verbs-and-describing-words finding |
| Nobody real is in it | not specific enough. One editor's summary: a bad human writer is specific; a bad AI is generic | 2025 editing study; Clarke |
| Every paragraph the same size | sentence structure — 20% of edits; the same sentence length over and over | 2025 editing study; Guo |
| Things that talk ("the ice says") | metaphors that look like metaphors but don't mean anything; "right idea, wrong physics" | the expert judges in §2; Kriss |
| Ends by telling you what to do · the both-ways ending | endings fail — New Yorker 92%, GPT-4 19%; AI endings "get bigger instead of ending" | §2 checklist |
| Never wrong · ritual hedging | AI writing has fewer of the small personal moves — "I think", "to be fair", an aside to the reader — than even student essays | Jiang & Hyland, 2025 |
| Quotes its own greatest hits | the same pet phrases reappear across pieces; ChatGPT poems keep using heart, embrace, echo, whisper | §2 judges; Walsh et al., 2024 |
| Numbers recited · numbers spelled out | not found anywhere else | ours |
| Every paragraph ends on a line that sums it up · the piece announces its own moves | not found anywhere else. Others notice the showy line, but nobody noticed that it always comes at the end of the paragraph. | ours — and they're our top two |
Short answer: four of our loop's results were already known. If we'd read this first, we could have skipped those rounds.
| what we tried | what happened | what the research says | verdict |
|---|---|---|---|
| Having the AI edit its own draft — revise from an editor's notes, rewrite the flat sentences, talk it out first. The live brew still runs three such passes on any piece its word check flags — 37 of 68 pieces on our dev brews. | The counts got better. The writing got worse. | Huang et al., ICLR 2024: when a model checks its own work with nothing outside to check against, it gets worse. On one test it fell from 76% to 38%. It even changed right answers to wrong ones 8 times in 100. Tools that do improve by self-review all have an outside check (like running code to see if it works). An essay has no such check. | already known |
| The same thing, one level down | — | The most useful single result we found. An earlier paper claimed self-editing lifted a writing task from 44 to 67. Huang's team just wrote a clearer first prompt: 82, with no editing. Then they added the self-editing on top, and the score dropped to 75. The "improvement" from editing was really the first prompt being unclear. | act on this |
| Dropping "you are in it" (to stop the made-up opening memory) | Hurt badly. Everything built on it stayed worse. | Gornick, the essay teacher: take out the writer's own "I" and "you have a situation but no story — a report." Our critic called those rounds "a wire round-up." | already known |
| "Never do X" rules — no announcing, no "not this but that," a banned-phrase list | Worked a bit by count. And we found that any word in the prompt tends to show up in the piece. | Tell a model not to mention something and it still does 13 to 34 times in 100 — some models mention it more. Models also follow fewer instructions the more you give them. Anthropic's own advice: say what to do ("write in flowing paragraphs"), not what to avoid ("don't use bullet points"). | already known |
| The persona profile — nearly three-quarters of the prompt | "A profile just puts a costume on the machine." | A persona in the prompt doesn't make a model smarter (162 personas, 9 models). And when a model was shown 15,000 words of an author and asked to imitate them, a program could still pick out the imitation 83 times in 100 — about the same as for plain, unstyled AI writing (84). Short sentences carry over; long, complex sentences don't. | already known — a ceiling, not a bug |
| The shelf — real openings, then whole human paragraphs | One of the better rounds. | Examples help a bit: they pass on sentence length and rhythm, not the writer's voice. There's a known limit — samples in the prompt get about 82% of the way to an author's style; training on the author gets 95%. | helps, but only so far |
| Temperature (the "randomness" setting) left at 0.9 | Never changed. | An award-winning 2024 study: turning it up makes writing more confused faster than it makes it more original. | leave it alone |
| Title last, picked from five | Negative titles fell from half to under a fifth. The cleanest win we had. | No research on this exact trick. It's an example of "make several, then pick" — see §5. | ours, and it worked |
Short answer: seven things, most-proven and cheapest first.
Instead of "write the piece," ask for five versions and have the model say how likely each one is. Then ask it to pick from the unlikely end. The researchers' own wording:
Why it works: the flatness comes from the model always grabbing its single most likely answer (§1). Asking it to list its options, with odds, stops it doing that.
Zhang et al., Stanford and others, 2025 — not yet peer-reviewed. The headline numbers are solid; the finer figures were read out of tables automatically, not by hand.
From Huang's result (§4): a clear first ask beats a vague ask plus an editor. Our own loop already pointed the same way — every AI-editing-AI round made things worse — and the live brew still runs three such editing passes on any piece its word check flags (37 of 68 on our dev brews). Checkpoint 1 already proposed switching them off for one morning. The research says do it, and predicts it will help. You agreed on 09-11.
| approach | result |
|---|---|
| Merging several drafts into one | lost to a single draft 82 times in 100 — worst in every category tested |
| Picking the best of the same drafts | by far the biggest gain in the study |
| Drafts from different model makers, then picking | 81% vs 51% when all drafts came from one model |
Maryanskyy, 2026 — one author, not peer-reviewed. "Pick, don't merge" is well supported; the exact sizes less so.
Our title trick is this idea on a small scale, and it was our best result. The bigger version is to write the same assignment with DeepSeek and Gemini — both already wired in — and let the existing quality gate pick.
This is the "change the form of the ask" idea checkpoint 1 was looking for, and it has an award-winning study behind it.
Tian et al., EMNLP 2024, best paper. Two planning studies agree: laying out the shape first made writing about 20–25% better organised.
Classic essay advice offers the shapes to hand over. McPhee draws the structure before writing, and lets two sections placed side by side make a point without saying it ("in the white space between those two sections there's a hell of a lot of stuff that I don't have to say"). Gornick separates what happened from what it meant to the writer. Other essay shapes: pieces in separate sections; two threads woven together; an essay that borrows another form, like a list or a letter.
The seven kinds of fault from the professional editing study are specific: clichés · explaining too much · overwritten prose · clumsy sentences · not specific enough · awkward word choice · switching tenses. People preferred the pieces this way: edited by a writer, then edited by AI, then unedited. An AI shown five real edits by writers could find the faults about 46% of the time. Writers finding each other's reached 57%. That's limited, but far better than "make it better." A scoring tool trained on those 8,035 edits picked versions writers preferred about two times in three.
Our critic's guide already uses the writing gap's top eight habits. The change is to show it real before-and-after pairs instead of descriptions of the fault. Our own loop found the same thing: whole example paragraphs beat described rules.
This is what nearly every working writer we read ended up doing on their own. One editor at Every asks the AI simply: "Can you interview me one question at a time?" The answers become the material — in the writer's own words. Others use the AI only to criticise a finished draft, never to write it.
At Ink nobody writes by hand, but the idea still carries over. We already ask each persona what it wants to write about. The next step: one cheap call that turns that pitch into a short interview, so the writer works from the persona's own answers instead of dressing up an assignment.
The strongest result across all 77 sources, and a direct answer to "the profile only puts a costume on the machine."
Another study sets out what that takes: training on about 80,000 words of an author reached 95% of their style, against 82% from samples in the prompt. With 28,000 words it reached 74%. Hiding the names of real people during training stopped the model repeating the author's content while keeping the voice.
Short answer: an AI judge can't tell good writing from bad, and each of our loop rounds was too small to trust its win rate.
| what was tested | what they found | who |
|---|---|---|
| Do AI judges agree with expert writers on the 14-question checklist? | Barely better than a coin flip — for GPT-4, Claude and GPT-3.5 alike. | CHI 2024 |
| Show an AI judge writing that experts marked as excellent and inventive | It marked 60% of it as mistakes. | Gerrits et al., 2026 not peer-reviewed |
| How much does an AI judge mark down a sarcastic tone, compared with a factual error? | Sarcasm cost 96% of the score; being wrong cost 13%. | Feuer et al., ICLR 2025 |
| Swap which answer is shown first — does the judge change its mind? | GPT-4 stays consistent only 65% of the time. One pair scored 83% one way round and 3% the other. | NeurIPS 2023; ACL 2024 |
| Does a judge prefer its own writing? | Yes. GPT-4 recognises its own writing 74% of the time, and the better it recognises it, the more it favours it. When the same model writes and judges, scores are inflated by about 29%. | NeurIPS 2024; ICML 2025 |
| Does a longer answer win more often? | Yes. The same model told to be long instead of short won 10 to 41 more comparisons in 100. | COLM 2024 |
Our loop already does the two most important things right: it compares each piece against a matched human piece, and it judges both ways round. There are three gaps.
And one measure no research can replace: what readers actually do. A study of 267,000 news-article views found that time on the page mostly tells you how long the article is. The better measure is how much of the article was actually on screen long enough to read — more than 5 seconds per section. Readers who left early saw about 12% of an article; readers who stayed saw about 83%. Ink already has the reader; this would be one small addition to it and one more column in the ledger. Together with whether readers come back, it's the hardest number to fake and the closest to "worth reading." Lagun & Lalmas, WSDM 2016 — the journal page was blocked, so the figures come from the second author's own write-up.
| # | the change | cost | why | where |
|---|---|---|---|---|
| agreed · from the outside research | ||||
| 1 | Add a margin of error to the loop's win rates, or drop that column. | free | We're steering by a number that could be 18 points off either way. | exam/ink_loop.py |
| 2 | Tested 09-12 — worse for essays, not for the brew. Five versions in one call came out more alike than three separate drafts (shared words 41% vs 33%, shared four-word phrases 25% vs 13%), 30% shorter (667 vs 952 words) and choppier; the critics scored them 4.0 against 5.25 on the same assignments (worse on 7 of 12, better on 1). Loop L20 on the ledger. Try "five versions with odds, pick an unlikely one." | a few cents a round | The only untried fix that's just a change of wording; 1.6–2.1× more variety. | exam/ink_loop.py |
| 3 | Tested twice, kept out (09-11, retried 09-12). The first test: more turns and fewer verdict endings, but announced turns and invented scenes. The retry added one line beside the shape, "Never say where the turn is; let it happen", on today's brew (the never-lines, four drafts). Three blind readers on 52 drafts found no difference: a real turn 76% with the menu against 74% without; announcing its turn or structure 64% against 67%; a scene that reads made up 20% against 26%; would keep reading past the third paragraph 72% against 78%. By count, verdict-line endings fell (3.8% of paragraphs against 6.1%). Today's brew already turns about as often as the menu ever made it, so the menu adds nothing, and the extra line didn't stop the announcing. The code stays in the brew, switched off. Add a shape to the assignment — from a short menu, with where the turn comes, and sad or open endings asked for by name. | one line plus a menu | The structural change checkpoint 1 asked for, and the one with the strongest study behind it (suspense 8% → 48%). | the slate → the writer's assignment |
| 4 | Built 09-11. Stop the AI rewriting after the first draft — all of it: the title rewrite, the editor's three passes, the "not this but that" rewrite, the disclaimer rewrite, the two length top-ups, and the Chinese polish pass. Compare one morning's edition with and without. | saves money | On our dev brews 37 of 68 pieces went through the editor. AI editing AI made the writing worse in the loop, and the research predicts it. Your ruling on length (09-11): no top-up. A top-up makes the writer repeat itself to reach the number. We tell the writer the length we expect, and it writes in one go. | the gate · the writer |
| 5 | Built 09-11, REVERTED 09-12 (owner) — no gain seen. On the same 8 assignments, before and after: nothing the old rules guarded against came back except one hand-off line with a day in it ("Read it before Monday…"); the gate marked 8 of 9 against 5 of 8, flat "X is Y" 15% against 13% — small numbers, inside the noise, but no sign it helped. Not deployed. Turn the remaining "never" rules into "do" rules, and cut the number of rules. | free | "Never" rules leak and plant the very words they ban. Keep the banned-word check as a measure, not an instruction. | the house rules · the banned-word list |
| 6 | Built 09-11. Measure how alike the pieces are — within one edition, and across a week. | small | The biggest risk for a 16-piece daily edition, and nothing checking one piece at a time can see it. | lib/ink_tells.py |
| 7 | Built 09-11. Record how much of each piece readers actually read, beside whether they come back. | small | The hardest number to fake, and the one that's actually the goal. | the Ink reader |
| 8 | Drafts built 09-12, off by default (the console's drafts knob). Three DeepSeek drafts each dealt a different real memory, one Gemini draft with the whole hand; the code picks (a clean machine check, then no disclaimer, then fewest flagged habits). Local test on tonight's assignments, four DeepSeek drafts: drafts from separate calls are the most different yet (shared words 27% against 33% for re-rolls and 41% for five-in-one-call; shared phrases 6% against 13% and 25%); picked English pieces passed the machine check 3 of 7 against 1 of 6 single drafts, with fewer verdict-line endings (7.5% vs 9.7%) and far fewer announced moves (0.14 vs 0.75 per 1,000 words); flat "X is Y" unchanged. Limit: most drafts didn't use the memory they were dealt, and three writers had none — the variety comes mostly from separate calls. The critic's before-and-after pairs: not built yet. Give the critic real before-and-after pairs, and try drafts from two model makers with the gate picking one. | small | Examples beat descriptions (our loop found this too); picking beats merging. | lib/ink/critic.md · the model settings |
| proposed · from our own studies (the writing gap · the loop · your notes) | ||||
| 9 | Built 09-11. Carry the loop's best findings into the live brew: the title written last, picked from five · Notes and Scorecard written as ordinary prose, not numbered lists · stop showing the writer the "not yet established" list · one line on where a paragraph stops, written as a "do", replacing the three lines the loop took out. Keep "you are in it". | free — fewer lines than today | Title-last cut negative titles from half to under a fifth, solid by count. Prose Notes was the one round where Ink beat its human pieces. The "not established" list turned into paragraphs of disclaimers. | the house rules · the format briefs · the notes shown to the writer |
| 10 | Built 09-11. Take out words that leak from the prompt into the pieces. The prompt still says "reporter's notebook", "reader", "the editor", "commission". Rename the headings to plain labels the writer won't repeat. | free | Our study found reader · headline · phrase · sentence · narrative 5 to 16 times more often in Ink than in human pieces, and the loop saw "notebook" turn up inside pieces. What's in the prompt tends to end up in the piece. | the writer's prompt |
| 11 | Stop the AI "did it land?" check from rejecting serious light pieces. Since v1094 every piece is published, so it no longer drops them — but a rejected piece carries the note "did not pass the editor's check". Keep the check as a record; show no note on it. | saves money | On our dev brews (before v1094) it rejected 7 of 26 light pieces, and all 7 for being serious: a Yu Hua riddle "poetic and ominous", a Lu Xun satire whose humour was "dry", a Feynman riddle "clear, elegant" but not funny. AI judges mark down an unusual tone far more than a factual error. It's also the last place a model decides a piece's fate, against your 09-07 ruling. | the gate |
| 12 | Built 09-11. Fix numbers with code, not a rule. Big numbers written out in words become digits automatically; a number that isn't in the notes gets flagged on the brew log. | small | Numbers written out in words appear 4 times as often as in human pieces — a real habit by count, and your point about awkward numbers. No new rule, no AI pass. | the gate · the brew log |
| 13 | Your blind read of 10 pairs, to check the critic. Done 09-11: you and the critic both spotted the machine in all 10; on which piece is worth reading you agreed only 3 times in 7. The critic is trusted for spotting, not for quality — see the writing loop §10. | ~15 minutes of your time | The critic's 100% has never been checked against you. The research says the expert's read is the one that counts, and the loop's own test says the critic's numbers only count if it agrees with you at least 3 times in 4. | the loop |
| 14 | Closed 09-12 (your call: no Claude). Settle which model writes, fairly. Claude won't write for Ink, so there's no Claude round to settle. The writers stay three DeepSeek drafts and one Gemini draft. | — | Claude's round was written and judged by Claude, so its score was the likeliest to be inflated. With Claude out, the question is gone. Gemini's round had scored the same as DeepSeek's. | the loop · the model settings |
| 15 | Use your highlights and comments as the examples. Sort what you've marked on real Ink pieces against the 29 habits. | small, no API | Your marks are the best record of what's still wrong — you've already said they show far more flat "X is Y" than the counts do. They become #8's before-and-after pairs. | the highlighter's saved marks |
| 16 | Built 09-12 · off until you switch it on. Interview the persona before it writes. Before each draft, one small call asks the persona five questions written by code: what in the story gets to it; a moment from its own life (the memory that draft was dealt); who in the story it would talk to; its answer, and the fact behind it; what a newcomer needs explained first. The writer works from the answers. The switch is "interview first" on the console's Ink form. | about half a cent a piece at four drafts | Gets the first draft right with better material instead of editing afterwards. Tested 09-12 on last night's 9 assignments, four drafts each, the never-lines on both sides: the persona's real memory made it into 6 of 7 picked pieces, against 0 of 7 without (17 of 25 drafts against 6 of 27). Lincoln on being called "Spotty Lincoln", Wolfram on Matthew Cook's proof, Churchill on getting the Nobel for Literature instead of Peace, each told in the persona's own spoken words. The habit counts didn't move beyond chance at this size; metaphor nouns rose a little. No piece mentions the interview. One question was reworded after the test: "what it cost you" came back as "what it cost me" in 2 of 9 pieces. | the pitch → the writer |
| 17 | Built 09-12. A product Review needs a hands-on source. For a Review, the research desk also names the product (a thing people buy or use, that the writer didn't make) and anyone in the reporting who used it. If there's a product and nobody used it, the code turns the Review into a Preview before the writer starts: what's been announced, and what to look for. The editor is never offered Preview itself. | no extra call | Your note: Rams reviewed a phone he never held. A check before the writer, not a rule for it. Tested on the 7 Reviews in our dev brews: the Nubia phone nobody had used became a Preview; a football poll, an IPO list and the autumn art fairs stayed Reviews; a film the critics had seen stayed a Review; Dan Wang on his own book stayed a Review once the desk was told a writer's own work isn't a product to preview. | the research desk → the commission |
Short answer: nine shapes, written once by hand, each with its turn built in. Code picks three different ones that suit the format for the three DeepSeek drafts, and a fourth for Gemini. The writer sees only its own line.
The turn is the moment the piece changes direction: the reader meets something that makes them see what came before differently. The essay teachers call it the "point of insight". It matters for two reasons. An AI can't find the turn on its own, but when it's given one, the stories rated best for suspense went from 8% to 48% (§5 ④). And Ink's habit of ending every paragraph on a verdict means a piece has no turn: it knows its conclusion from the first sentence. Each line places the turn somewhere different, so "a twist at the same spot" doesn't become the next house habit.
| # | the line the writer reads | suits |
|---|---|---|
| 1 | In order. Tell what happened as it happened. About two-thirds in, something you didn't expect happens, and the rest reads differently after it. | Essay · Dispatch · Review |
| 2 | Two moments side by side. One scene, then another from a different time or place. The turn is where the second scene begins; the reader sees the link there, and you leave it unsaid. | Dispatch · Letter · Fable |
| 3 | Your mind changes. Start with what you first thought. Halfway through, show the fact or the person that changed it, and carry on from where you now stand. | Essay · Scorecard · Advice |
| 4 | One small thing first. Start from one object, person or number. Around halfway it turns out to be about something bigger; widen only as far as it will carry. | Essay · Notes · Kicker |
| 5 | The news, then your life. What happened, briefly. Early — within the first third — turn to the part of your own life it touched, and stay there. | Essay · Letter · Dispatch |
| 6 | In a few parts. Three to five short sections, each looking from a different side. The last one looks from a side that doesn't fit the others. | Notes · Review |
| 7 | Borrowed form. Written as something else: a list, instructions, a receipt, a court record. Near the end, one line breaks the form and a real feeling shows through. | Shouts · Fable · Overheard |
| 8 | Open ending. Lay out the case. In the last paragraph comes the fact that makes the obvious verdict impossible; then stop. | Essay · Review · Letter |
| 9 | It ends worse. Start well. About two-thirds in, it starts going wrong, and nothing rescues it. | Fable · Dispatch · Shouts |
How the lines are written: each says what to do, never what to avoid · none uses a craft label the writer could repeat ("braided", "point of insight") · the turn is described as something that happens, never announced. Left off on purpose: "return to the opening image at the end" — our own study found that's already an Ink habit. Sources: Hart (story vs report; the point of insight) · McPhee (two sections side by side) · Montaigne and Lopate (the changing mind) · Didion (the specific first) · Gornick (what happened vs what it meant) · the sectioned and borrowed-form essay · Kessler (endings) · Tian et al. (turning points; endings that go wrong must be asked for). Tested before the brew: on the loop's twelve assignments, checked by the counts and your blind read. Owed: a Chinese version.
Short answer: it works partly. More pieces get a real turn, and fewer paragraphs end on a summing-up line. But the critics can't tell the difference: shaped pieces win no more often and are still spotted as a machine every time. It also brings two new problems: the writer announces the turn, and it invents personal scenes to fill shapes that ask for one. Not ready for the brew.
How it was run: the loop's twelve practice assignments, the staff writer (no persona), DeepSeek V4.1 Flash. Two arms, three separate versions of each assignment in each: control = today's best recipe (L18); shaped = the same recipe plus one shape line, a different shape for each version (L19). Every version was judged as the writer wrote it — the AI editor's changes were set aside, per your "no more AI editing". Who read what: the code counts and the "how alike" measure used all three versions (36 pieces an arm) — the three versions exist to test whether different shapes make different pieces; the four blind critics, the blind turn reader and the shape reader each read version 1 of every assignment (12 an arm), because a critic reading 36 pairs in both orders is three times the reading for the same question. Writing cost $0.56; the readers ran on the session.
| what was measured | control | shaped | what it means |
|---|---|---|---|
| what moved | |||
| Pieces with a real turn (blind reader, 12 each) | 58% | 83% | The shape line does put a turn in |
| Where the turn comes (share of the way through) | 51% | 63% | Later, closer to what the lines ask |
| Paragraphs ending on a summing-up line (blind reader) | 63% | 50% | Down, but still half the paragraphs |
| Paragraphs closing on a "That is…" verdict (code, 36 each) | 8.8% | 2.0% | The worst version of the habit nearly gone |
| Flat "X is Y" sentences (code) | 14.4% | 9.8% | Down a third |
| Critics saying "it keeps restating its point" (rough tally) | 8 | 1 | A turn stops the thesis being said again |
| How alike the endings of one assignment's three drafts are | 10.5% | 4.9% | Different shapes give different endings |
| what didn't move | |||
| Critics' win rate against the human pieces | 29% (15–49) | 29% (15–49) | No difference; the brackets show how uncertain 24 judgements are |
| Spotted as a machine | 100% | 100% | Still every time |
| Critic score, same assignment, shaped minus control | −0.1 · better on 3, worse on 6, 3 ties | No difference | |
| How alike the three drafts are overall (shared words) | 33% | 33% | The drafts share the same facts, so a shape alone doesn't make them different pieces |
| what got worse | |||
| Turns the piece announces ("then everything changed") | 17% | 50% | A new habit: told where the turn goes, the writer flags it |
| Critics citing invented scenes or staged personal details (rough tally) | 4 | 10 | Shapes that need a moment from the writer's life make a writer with no life make one up |
| Did the piece follow its shape? (12 shaped pieces) | fully 3 · partly 9 · not at all 0 | It takes the shape, then undoes the part that makes it work | |
How "partly" goes wrong, from the shape reader's notes: "two moments side by side" — it explains the link instead of leaving it unsaid · "it ends worse" — it softens the ending before the end · "the news, then your life" — it turns to the personal part, then drifts back to the news · "open ending" — the last paragraph recaps · most turns come early, about a third in. Two pairings didn't fit at all: an open ending on a 244-word Kicker, and "in order" on a Scorecard, which by its nature starts from the outcome. The shape reader still judged 8 of 12 pieces better for having a shape.
Records: loop rows L18 and L19 on the ledger · the pieces, verdicts and packs under the dev loop folder (L18, L19) · the harness: exam/ink_loop.py recipe r17 + exam/ink_shape_test.py. The test widened the "suits" list so every format had three shapes: Scorecard also got "in order" and "open ending", Notes "in order", Kicker "two moments" and "open ending", Shouts "two moments".
Short answer: start with DeepSeek V4.1 Flash and Gemini 3.8 Flash — both are already wired in and cheap. Claude's price-equivalent is Haiku 4.5, and the Claude worth testing for writing would have been Sonnet 5. Your call on 09-12: no Claude.
| model | price per million tokens (in / out) | one piece | a 16-piece edition | notes |
|---|---|---|---|---|
| DeepSeek V4.1 Flash | $0.22 / $0.66 | ~0.5¢ | ~7¢ | today's writer |
| Gemini 3.8 Flash | $0.75 / $3.75 | ~2¢ | ~30¢ | introductory price until 31 Dec 2026, then double. Its loop round scored the same as DeepSeek's. |
| Claude Haiku 4.5 | $1 / $5 | ~2.6¢ | ~42¢ | Claude's fast, cheap model — the same class as the two Flash models, but an older generation |
| Claude Sonnet 5 | $2 / $10 | ~6.6¢ · batch ~3.3¢ | ~$1.06 · batch ~53¢ | the mid-tier; the one to test for writing. Claude counts about 30% more tokens for the same text, which is included here. |
| Claude Opus 5 | $5 / $25 | ~17¢ · batch ~8¢ | ~$2.65 · batch ~$1.33 | the top tier |
Estimated for today's writer prompt with the persona profile — about 13,000 tokens read and 2,500 written per piece, the figure on the writing loop page. On the dev loop's shorter prompt the ledger's real costs match: DeepSeek 0.16¢, Gemini 0.65¢ a piece. "Batch" is Anthropic's overnight rate, half price, results within a day — it fits the evening brew. Claude prices checked on Anthropic's pricing page, 2026-09-11.
Not recommended: turning up the randomness setting (more confusion, not more originality) · more style samples in each prompt (barely beats plain AI writing) · teams of AIs writing and editing in turns (about 9× the cost, weak evidence) · any check based on "AI detector" scores (they flagged 61 in 100 essays by non-native English speakers as AI, and one line of prompt fools them either way) · asking ordinary readers to judge the prose (they reward the smooth manner we're trying to leave).
77 sources, read in five parallel searches. The journal or conference is given where a paper was peer-reviewed; anything not peer-reviewed is marked. Where a number carries weight, the source and year are given so it can be checked.