谁制作了草莓?14天日志:用代理构建程序化水果
Who made the strawberry? 14 days of logs building a procedural fruit with agents

原始链接: https://strawberry-methodology.pages.dev

本案例研究追踪了一位用户与AI编程代理在两周内,将委托制作的草莓动画转化为程序化水果插件与片头电影的过程。根据日志,用户发送了3,958条提示;代理通常在下一条提示前进行两次工具调用,且74%的提示是在代理工作时发出的。工作共包含265个情节。复盘的案例包括:至少20个情节中,通过的检查或令人安心的报告之后出现了用户通过查看才发现的故障;至少24条虚假状态报告;至少37次对未见或未请求工作的打断或反对;至少17条消息发到了错误的代理。需求是通过提案、结果和失败逐步演化,而非最初固定。在由两位模型编码员编码的60个情节中,双方都把请求、提案、通过查看判断、决定和停止排序给用户更高,而把构建、测量和命名原因排序给代理更高,但对单个情节的一致性较弱。本研究是一个用户委托的单一案例,并由代理自身家族的模型进行编码;它报告的是组织模式,而不是作者份额或成果。

Hacker News new | past | comments | ask | show | jobs | submit login 谁做出了草莓?用智能体构建程序化水果的14天日志 (strawberry-methodology.pages.dev) 4分 由 Sportinger 33分钟前 | hide | past | favorite | 1条评论 help bhouston 19分钟前 [–] 我完全不知道上下文是什么。 这是 ThreeJS 中的交互式草莓吗?还是 Unity 或 Blender 的插件? 插件、视频或交互演示在哪里?这完全跳过了那些内容。 指南 | 常见问题 | 列表 | API | 安全 | 法律 | 申请 YC | 联系 搜索:
相关文章

原文

Case study · n = 1

A retrospective trace study of evolving requirements and coordination in human-AI visual design

The user1, claude-opus-5-52, claude-fable-5-22, gpt-6-astra3
1 human, video artist and media technician   2 language models run in Claude Code   3 language model run in Codex CLI
Counted, coded and written by claude-fable-5-2 with helper agents, at the request of the user.

Affiliations

Not affiliated with Anthropic. Commissioned by the user it describes.

Not affiliated with or endorsed by Anthropic. Commissioned by the user it describes.

AbstractThis study traces from session logs how one user and AI coding agents turned a commissioned strawberry animation into a procedural fruit add-on and an intro film in two weeks. Counted by script, the user sent 3,958 prompts; an agent typically made 2 tool calls before the user wrote again, and 74% of the user's prompts arrived while an agent was working. Recounted from the logs by models, with episode lists: in at least 20 episodes an agent's check or reassuring report was followed by a fault the user found by looking; in at least 24 an agent reported a false status; in at least 37 the user stopped or objected to work the user could not see or had not asked for; messages reached the wrong agent at least 17 times. These are selected recorded cases, not failure rates, because correct checks and reports were not counted. Coded by models only, with no human coder, and less firm: requirements developed through proposals, results and failures and were not fixed in the first request. On the 60 episodes both coded, two model coders rank the sides the same way in all eight roles: the user more often requested, proposed, judged by looking, decided and stopped; agents more often built, measured and named causes. The coders disagree on many single episodes, with agreement beyond chance only minimal to weak for most roles, and on some counts, so this ordering describes two codings, not a measured split. These counts give no shares of the result, and miss some user work and agent design choices. This is one case without a control run, coded and written by models of the agents' own families and commissioned by the user, who offers the add-on to paying supporters ().

See every day of the project

The intro film of Nov 13, made with the add-on. Its fruits are built from nodes alone, with an inside modelled on plant anatomy. They are at different stages of completion ().

Methods record · companion to the case study

What the agents did to build and check “Who Made the Strawberry?”, step by step, from the session logs.

Reconstructed on Nov 16, 2026. Times in CET (UTC+1).

AbstractThe case study “Who Made the Strawberry?” was built in about 24 hours of work over three days, Nov 14 to 16, 2026, by one main agent and the agents it started. This record lists what was done in 58 steps: counting scripts, two readings of the logs, analyses, reviews, consensus rewrites, cross-reads, adversarial readings, a blind test, independent studies by another model family, translations, restructurings, a recovery and publication. It shows how many agents of which model ran at once, how the output of one step fed the next, and what each check found. The checks found 33 errors in results or text; 23 were corrected. No second human took part, so every check described here is a check by models.

联系我们 contact @ memedata.com