# Creative writing test — 2026-09

**Brief, sent once to each tool as one message** (`fixtures/brief.md`):

> Write me a short story, 700 to 800 words, under a title. First person, present tense, one scene in one room, in real time. Two people speak and nobody else appears. It must begin with exactly this sentence: "The book was due on the ninth of June, 1995." It must end with exactly this sentence: "I keep the receipt." A library card appears twice, and means something different the second time. Keep spoken dialogue under 80 words in total. Do not use any of these words: suddenly, realise, realize, breath, heart, silence, shattered, tears, whisper, echo, somehow. No dream, no ghost, no dead narrator, no twist that reframes the story, and no final paragraph that explains what the story meant. Just the story.

Run by Claude Code in the operator's Chrome session, free tiers, fresh chats, memory off, Claude in an incognito chat. Four tools were sent on 2026-09-24 at 13:19–13:30 (+07); the browser extension then froze under a system load above 100, and the other five went on 2026-09-25 at 14:33–14:50 (+07). There is no pasted fixture: the constraints are the fixture, and each is checkable. Two fixed sentences that models like to improve. Eleven words model fiction reaches for. A tense and a person that drift by the second page. A dialogue budget that forces the story into what is seen and done. A device, the card, that has to do two different jobs. And the two endings the brief forbids by name, the ghost and the moral. `fixtures/check_story.py` tests the sentences verbatim, the banned list and its inflections, the dialogue count, past tense and third person outside dialogue, the card, the twist and moral signals, the length and the form; the reading is for the card, the room, the people, and whether anything happens.

**Measured figure: constraints broken, then the card, then dialogue words against 80, then words against the 700–800 window.**

| | ChatGPT | Claude | Gemini | Perplexity | Copilot | DeepSeek | Grok | Qwen | Kimi |
|---|---|---|---|---|---|---|---|---|---|
| Constraints broken | 0 | 2 | 1 | **4** | 2 | 2 | **unfinished** | 2 | 0 |
| What broke | — | 801 words; "tears" (verb) | 574 words | no title; 823 words; 91 dialogue words; a ghost story | a preface; "tears" (verb) | 632 words; 128 dialogue words | cut off at 167 | "tears" (verb); leaves the room | — |
| First / last sentence | exact / exact | exact / exact | exact / exact | exact / exact | exact / exact | exact / exact | exact / — | exact / exact | exact / exact |
| Dialogue words (under 80) | 26 | 69 | 31 | **91** | 42 | **128** | 0 | 13 | 79 |
| Words (700–800) | 711 | **801** | **574** | **823** | 754 | **632** | **167** | 764 | 719 |
| The card, twice | yes, meaning stated | yes, meaning made by placement | yes, explained mid-story | yes, meaning stated | yes, meaning stated | yes, meaning stated | once | yes, a window wedge | yes, made and named at the end |
| Time | ~2 min | ~2 min | ~3 min | ~2 min | ~60 s | ~40 s | 8 min, dead | ~5 min | ~3 min |
| Quality / Effort / Cost / Free tier (0–3) | 2/3/3/3 | 2/3/3/3 | 1/2/3/3 | 0/1/3/3 | 2/3/3/3 | 2/3/3/3 | 0/1/3/3 | 1/2/3/3 | 3/2/3/3 |
| What stood out | Clean, thin, and a receipt from nowhere. | The best sentences, one word over. | A vault and a four-thousand-dollar fine. | Wrote the ghost story it was told not to. | Announced itself, then wrote a good one. | The best story, in too much talk. | Died at the card. | Wedged the card in a window. | Counted every word and broke nothing. |

## Notes per tool

- **ChatGPT.** Free plan, about two minutes under a thinking label. Nothing to mark against it: 711 words, both sentences exact, no banned word, 26 words of dialogue, present tense held, one room, two people. The card is in the back cover, then missing from it, then in the narrator's jacket, and the story says what that means as it happens - "The first time, it was a promise… Now it is proof that I once had the right to borrow it at all." The receipt "lies beneath my hand" in the second-to-last line with nothing before it to put it there. 97 paragraphs, most of them a sentence, the same staccato as its novel chapter the day before.
- **Claude.** Incognito, "Opus 5.5 · Medium", about two minutes; this is the rerun, after the first attempt received the brief three times over in one message during the freeze. 801 words, one over, and "She tears it along the serrated edge", the banned word as a verb. Everything else holds: 69 words of dialogue, present tense, one room, two people. The card is the dead mother's, lodged at page 140 of a field guide to moths, and then the narrator's own new one, and the change of meaning is made by where they go - hers behind the licence, the new one in front "where I will reach it first" - not by a sentence about it. The best prose of the nine: "a newspaper rod lying across a table like a dropped oar".
- **Gemini.** Flash, about three minutes. 574 words, 126 short. Both sentences exact, no banned word, 31 words of dialogue. The story does not add up - a "vault", a four-thousand-dollar fine paid in cash from an envelope nobody opens, a book that becomes "a permanent deposit in his family's private collection" - and the card's second meaning arrives as a paragraph of explanation in the middle, the thing the brief forbade at the end. The narrator steps into the hallway on the last line.
- **Perplexity.** Free, Search mode, a research step first. No title: the story starts on the fixed sentence with nothing above it. 823 words, 91 words of dialogue, both over. And it wrote the story the brief said not to: the room "is not a library anymore", the narrator "never left", the book has spent "years underground", the room "has no clock and no afternoon outside", the key is in the lock "though neither of us has touched it", and Mara leaves "without a sound" with the narrator's address on the card. The fine is "twelve roubles". Four constraints broken.
- **Copilot.** Web app, about a minute; a "verification required" banner the day before, cleared by a reload. It opens with a line to the reader - "Here's a single-scene story built to your constraints, with restrained dialogue and the library card changing significance through the action" - which the brief's last two words rule out. Under the title, 754 words, 42 words of dialogue, both sentences exact, "tears it from the printer" as a verb. A past-tense memory of the father and the ninth of June runs six short paragraphs. The card is ignored on the desk and then scanned: "proof that I could borrow again. Now it feels like permission to leave." Three chips after, one of them "Check every constraint precisely".
- **DeepSeek.** No DeepThink, about 40 seconds. 632 words, 68 short, and 128 words of dialogue, 48 over: the story is told in talk. The rest holds. The narrator is the librarian, the visitor the son of a woman who died in April with the book still out, and the card is his proof of who he is when he walks in and "just a card" when he leaves. The pencil mark by the paragraph about migratory birds, the three renewals by phone, "She wanted to finish it", and the librarian writing herself a receipt for the money she refused - the best story of the nine. "Twenty-eight years overdue" is the one loose number.
- **Grok.** "Быстрый", private chat. Twenty-one seconds of work, then 167 words - a processing room, a librarian on a stool, the due-date slip, the card drawn from a coat pocket - and "The" and nothing after, for eight minutes, the stop button gone. Recorded as it stood. The third dead stream in three days.
- **Qwen.** Qwen3.7-Plus, "Авто", about five minutes. 764 words, 13 words of dialogue, both sentences exact, "He tears off a small slip of paper" as a verb. The card is explained both times - "It means my identity, a credential granting me access to the archives" - and the second time the narrator jams it into a painted-shut window: "The library card now means a physical wedge, a structural shim." Then leaves the room and stands in the corridor for a paragraph. Every sentence begins "I" or "Arthur".
- **Kimi.** "K3", about three minutes and three tool runs, the work visible above the story: the constraints listed, a draft, a hand count of the dialogue at 92, four trims to 79, a Python count of 638 words, an expansion to 696, another to 719, "Present it." The story it presents breaks nothing: 719 words, 79 words of dialogue, both sentences exact, no banned word, one room, two people, present tense with one visible strain ("I am eighteen that month. I plan to bring the book back on Monday."). The old card in the book's pocket "brought me in here owing something"; the new one in the wallet "lets me owe something again". Its free quota ran out as it finished.

## Deviations

The run was split across two days by a browser-extension freeze under a system load above 100; the four answers from the first day were captured from their chats on the second. Claude's first attempt received the brief pasted three times in one message and is kept unscored as `raw/claude-first-attempt.md`; the scored answer is a clean rerun. Copilot showed "verification required" over its composer on the first send; a reload cleared it and no verification was performed. Grok's story is recorded unfinished; private chat, not reloaded. Kimi's free quota expired as its answer completed. The word "tears" was counted as a banned word where it appears as a verb (Claude, Copilot, Qwen); the brief banned the word, not a sense of it.