Observed facts: Simon Willison says he used the default Claude iPhone app in chat mode, not Claude Code, to request a Three.js game called Tree Appreciator. He reports that Claude Opus 5.5 Max worked for 3 hours and 7 minutes. The resulting game presents two algorithmically generated trees for players to compare and includes music, animation, and achievements.
Mechanism: The user gave a broad goal and quality direction, and the model spent an extended session generating and refining an interactive experience; the account does not reveal internal steps or whether a human intervened. This is the mechanism visible in the report: a broad goal and quality direction were followed by an extended generation session that produced an interactive experience, although internal steps and possible human intervention remain unreported.
Why now: Long-running work in a familiar consumer chat interface makes the boundary between conversational prompting and software-building less dependent on a dedicated coding environment.
First implication: Builders can test broad goal prompts for interactive prototyping while retaining human review and reproducibility as unresolved requirements. Builders can test whether broad goal prompts produce useful interactive prototypes, while treating human review and reproducibility as unresolved requirements.
Evidence boundary: This is one author's account of one experiment, with a linked playable tool and shared conversation; it supplies no controlled comparison, independent verification, repeatability rate, cost accounting, or account of review effort.
First watch condition: Watch for independent repetitions that report completion rate, human intervention, and review effort for similarly open-ended game-building tasks.
The experiment therefore illustrates a potentially useful interaction pattern, but its duration and polished feature list should not be mistaken for evidence of consistent autonomous performance. The next meaningful evidence would show whether another user can reproduce the outcome and how much correction is needed before the game is considered finished.