Case study · n = 1
Work in progress...
Who Made the Strawberry?
A case study of one human and three language models building a procedural fruit add-on for Blender.
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.
AbstractOne video artist and several AI agents built a procedural fruit add-on for Blender in two weeks. This case study reconstructs the project from its own traces: the prompts the user typed or dictated, counts of the agents' tool calls, screenshots and commits. It asks what the human contributed and what the models contributed. The agents built everything. The user started nearly every episode, gave the direction and judged the result by looking. In most episodes neither side decided the outcome alone. The study is one case, coded once by AI agents, at the request of the user it describes.
1Introduction
A common reaction to media made with the help of AI agents is that the human did nothing. This project is a good case to check that claim, and to see what the agents' part really was, because almost everything about it was logged: about 2,000 prompts, about 18,000 actions by the agents, and more than a thousand screenshots.
It began on October 31 as a freelance job: animate a strawberry for an ad. In the end there was no money for it. The user kept going on his own and set himself a new goal: a strawberry built only from nodes, with no downloaded model and no textures, correct on the inside when you cut it. Two weeks later there was an add-on with six fruits and an intro film.
This page was made the day after the result was posted, at the request of the user, to answer that reaction. It is not neutral ground. What it can offer is counting instead of arguing: who did what, where things went wrong, and a link from every claim to the episode it rests on. Its weak points are listed in section 4.1.
1.1Questions
- The analysis is being written.
2Methods
2.1Setup
The user. A video artist and media technician who has made video for theatre and opera since 2011. He studied biology for some years, then camera. He has written code for media work for years and has built software products with AI agents since 2025. In this project he wrote no code and built no node trees by hand.
The user described in plain language what he wanted, mostly by dictation, and sent screenshots of the Blender viewport. The agents wrote and ran Python inside Blender, looked at their own screenshots and reported back. The user worked in the same Blender file by hand: camera, light, materials, and looking.
Table 1. Models and harnesses. A harness is the program that runs a model and gives it tools. “Prompts” are the user's prompts answered by that model. “Log entries” are the model's replies, thoughts and tool calls.
Figure 1. Which model the user worked with on which day, by share of his prompts.
2.2Data and coding
- Sources. Session logs of both harnesses, the git history, the user's screenshots and the project folders. Nothing was recorded for this study. The logs already existed.
- Counting. A prompt is one message the user typed or dictated. Messages passed on from other agents and exact repeats on the same day are left out. A tool call is one action of an agent, for example running a script in Blender or reading a file. Status reports to the terminal are left out.
- Episodes. For each day, one agent read all prompts of that day and the first 500 characters of each agent reply, and split the day into episodes. Each episode got one label: went well, decision, routine, misunderstanding, dead end or stuck. The labels are a judgement, not a measurement.
- Analysis. A further agent read all episodes, looked for patterns that repeat and counted them. Each finding links to the episodes it rests on.
- Quotes. The user writes and dictates in German. Quotes are translated.
- Origin and interests. The study was started on November 14, about 14 hours after the result was posted on Reddit, to be linked from there. The user sells the add-on. Everything here was counted, coded and written by AI agents of the same model family as the ones described.
- Review. A further agent of the same model family reviewed the study and checked samples against the logs. Its corrections are built in. Its open points are listed under Limits.
3Results
3.1At a glance
Table 2. Totals counted from the logs, with the exclusions named in section 2.2.
prompts from the usertool calls by the agents
Figure 2. Activity per day. The two rows use different scales. Click a day to jump to it.
3.2Four stages
3.3Where the time went
Figure 3. Where the time went, as shares. Top: all days together. Below: each day, with its number of active hours on the right. Episodes overlap because sessions ran in parallel, so these are shares of episode time, not hours of work. Labels are the reading agent's judgement.
3.4Who decided the outcome
Figure 4. All 263 episodes, coded three ways. “Both” means the user asked for something or saw a symptom and an agent built it or found the cause.
Figure 5. Kinds of contribution, counted in how many of the 263 episodes each appears. One episode can hold several kinds.
Episodes the user decided
Episodes an agent decided
3.5How the user worked
3.6Findings
3.7Who did what
Table 3. Division of work between the user and the agents.
3.8Sound
3.9How the audience reacted
What the comments are about
Comments about AI: what did the human do?
What viewers noticed in the picture
4Discussion
4.1Limits
- Conflict of interest. This page was written by the same kind of models it describes, at the request of the user it describes. The counts can be checked against the linked episodes. The judgements cannot be fully neutral.
- Second hand. Episodes are summaries written by an agent. The analysis reads those summaries, not the raw logs.
5Conclusion
ADay-by-day record
Each bar is one day from 0:00 to 24:00. Each coloured block is one episode. Click a block or a row to read what happened.