Inside the Claude Code Session that Built a 3D Boeing 747
victor/fable-5-boeing-747-trace ↗
The **victor/fable-5-boeing-747-trace** dataset is a compact (under 1 KB) JSONL collection that records an entire Claude Code (Fable 5) agent session. Authored by "victor" and released under an MIT license, the dataset contains one JSON object per line representing user prompts, assistant messages, tool calls, and the resulting data—including base64‑encoded screenshots of the Three.js renders. It is tagged as an "agent‑traces" dataset and is compatible with the Hugging Face `datasets`, `pandas`, `polars`, and `mlcroissant` libraries.
The trace documents how the Claude Code agent built a procedural Boeing 747 model in Three.js from a single `/goal` prompt. Over roughly 30 minutes, the agent iteratively generated code, launched a headless screenshot camera, and entered a vision‑based self‑verification loop, refining the model until it satisfied a visual quality check. The session also includes two follow‑up goals: a 12‑second cinematic camera fly‑through and a full‑scale airport scene surrounding the aircraft.
Because the transcript captures both the textual reasoning and the visual feedback loop, it serves as a valuable resource for researchers studying LLM‑driven code generation, tool‑use planning, and multimodal self‑verification. The sanitized data removes any personal identifiers while preserving the full interaction flow, making it safe for analysis, replay, and educational purposes.
The dataset has quickly risen in popularity, reflected by a trending score of 16 and over a thousand downloads within days of release. Its uniqueness lies in exposing the end‑to‑end workflow of an AI agent that combines natural‑language reasoning, program synthesis, and computer‑vision verification to create a complex 3D asset, providing a rare glimpse into the inner workings of advanced coding agents.
Project Ideas
- Create an interactive replay viewer that parses the JSONL trace and sequentially renders the base64 screenshots to visualize the agent’s decision‑making process.
- Fine‑tune a small language model on the transcript to predict the next tool call or code snippet when given a similar 3D modeling goal.
- Analyze the self‑verification loop by extracting iteration counts and screenshot timestamps to study how often the agent revises its output before reaching satisfaction.
- Build a step‑by‑step tutorial generator that converts the transcript into markdown instructions for reproducing the Boeing 747 model in Three.js.
- Develop a benchmark suite that measures how well other coding agents can replicate the same 3D modeling task using this trace as a reference.