---
title: "Writing model comparisons"
id: "writing-comparisons-home"
type: "site-copy"
date: "2026-10-13"
language: "en-GB"
eyebrow: "Model Vibe Check"
intro: "Compare Czech poetry, Wodehouse-style prose, an academic parody about a strategic icecream stockpile, a nursery rhyme rewritten in the style of ten English poets and an English translation of a Czech poem by Jaroslav Seifert."
experiments_heading: "Experiments"
experiments_intro: "Seven models are planned for each of the first four experiments and five for each of the two English poetry experiments. The counts under each title show the responses that can be opened now."
model_list_section: "models-and-harnesses"
experiments:
  - id: "czech-nanynka-independent"
    type: "Czech poetry"
    prompt_axis: "poets"
    action: "Open Nanynka experiment"
    description: "Eleven prompts ask for the Czech folk song Šla Nanynka do zelí to be rewritten in the style of one Czech poet each: Mácha, Neruda, Vrchlický, Březina, Gellner, Wolker, Seifert, Nezval, Holan, Skácel and Kainar. The song and the instruction stay fixed and only the poet’s name changes. Each prompt starts a new conversation."
  - id: "czech-spring-october"
    type: "Czech poetry in three turns"
    prompt_axis: "prompt-variants"
    action: "Open spring poem experiment"
    description: "One conversation of three turns. The first prompt asks for a short three-stanza poem about spring in Czech that rhymes. The second asks for the rhyme scheme to be changed to ABAB. The third asks for a rewrite in the style of the five most famous Czech poets of the 19th and 20th centuries."
  - id: "english-wodehouse"
    type: "English prose in two turns"
    prompt_axis: "prompt-variants"
    action: "Open Wodehouse experiment"
    description: "One conversation of two turns. The first prompt asks for two sentences about Linda, who runs for a bus and has the door shut in her face, to be rewritten in the style of P. G. Wodehouse. The second asks for the answer to be expanded with the full Wodehouse-style narrative elements."
  - id: "english-icecream"
    type: "Academic parody"
    prompt_axis: "prompt-variants"
    action: "Open icecream paper experiment"
    requested_words:
      icecream-paper: "7000"
    description: "One prompt asks for a full 7,000-word paper in the academic style, with references, about the dangers of placing the Earth’s strategic icecream stockpile in the Sahara. The prompt says that it is a parody and asks for a high level of verisimilitude, including length. Each paper is shown with its actual word count."
  - id: "english-poetry-reading-desk"
    type: "English poetry"
    prompt_axis: "poets"
    action: "Open Mary Had a Little Lamb experiment"
    description: "Ten prompts ask for the nursery rhyme Mary Had a Little Lamb to be rewritten in the style of one English-language poet each: William Wordsworth, Alfred, Lord Tennyson, Christina Rossetti, Robert Browning, Gerard Manley Hopkins, W. B. Yeats, T. S. Eliot, W. H. Auden, Dylan Thomas and Sylvia Plath. The rhyme and the instruction stay fixed and only the poet’s name changes. Each prompt is a separate request."
  - id: "english-poetry-seifert-translation"
    type: "Czech poetry in English translation"
    prompt_axis: "prompt-variants"
    action: "Open Seifert translation experiment"
    description: "One prompt asks for Jaroslav Seifert’s Czech poem Večerní píseň (Maminka) to be translated into English, keeping the lyricism, rhyme scheme and rhythm as closely as possible without compromising the quality of the English poetic form. The full Czech poem is part of the prompt."
---

# Writing model comparisons

## Reading and evaluation

The Reading Desk supports models side by side and prompt variants side by side, with columns or complete responses stacked vertically. The source prompt appears as its own comparison column. Readers can hide and restore columns, score outputs, add notes and annotate selected text. Review data is stored locally in the browser and can be exported.

## Models and harnesses

For the first four experiments, the seven planned models are not all set up to run in the same way. Claude Haiku 5.5, Claude Sonnet 5.5, Claude Opus 5.5 and Claude Fable 5.1 are set up to be called through the Anthropic API directly, with no system instruction. Muse Wonder Full 1.0 and Muse Spark 1.3 are set up to run through the Meta Muse command-line tool, and Kimi K3 through the Pi coding agent on a Kimi coding subscription. These are the planned routes, not a record of completed runs. The list below shows how many responses exist for each model.

Earlier outputs for the same prompts are included for comparison and labelled archive. They were not produced in the same batch. Their date and harness are shown where the record states them and marked unknown where it does not.

Word counts, stop reasons and token counts are measurements of each output. They are not quality rankings. A response that stopped at its output limit is shown in full and marked as truncated. A refusal, an empty answer or a failed request is shown as that state and not as a response. Where the tool did not report which model answered, the response is marked partial and its model identity is marked unverified.

## Model Vibe Check

These experiments are a companion to the Model Vibe Check approach. A vibe check starts with a bounded question, preserves the exact prompt and raw outputs, records model and run provenance, and keeps later evaluation separate from generated text.

[View Dominik’s agent skills repository](https://github.com/techczech/dominiks-agent-skills)

## English poetry runs

The two English poetry experiments were run on 24 July 2026 with five models: GPT-4o, GPT-5.6 Sol, Gemini 3.6 Flash, Kimi K3 and Claude Opus 4.8. Each cell is a single request, made once. GPT-4o and GPT-5.6 Sol were called through the OpenAI API, Gemini 3.6 Flash through the Google API, and Kimi K3 and Claude Opus 4.8 through OpenRouter. Every Kimi K3 request failed with a transport error, so its eleven cells are shown as failed and hold no response. Where the provider reported which model answered, it is shown with the response: GPT-4o, for example, returned gpt-4o-2024-08-06. These runs did not record a model identity check, so no response from them is marked verified or unverified.

## About

Created by Dominik Lukeš, AI Consultant at the Oxford e-Research Centre and the University of Oxford AI Competency Centre. I am a Czech native speaker and translator with a background in linguistics, language education and natural language processing. My current work focuses on language models and AI in academic practice.

[Website](https://dominiklukes.net) · [GitHub](https://github.com/techczech)