This dataset exposes a compact, real-world example of a multimodal agent execution: a Claude Code (Fable 5) session that iteratively built a procedural Boeing 747 in Three.js using a vision‑enabled self‑verification loop.
What Sets It Apart
- Full-session JSONL transcript including user prompts, assistant thinking, tool calls, tool outputs, and embedded base64 screenshots — not just high‑level logs. This makes it possible to replay or analyze the agent’s stepwise decisions and visual checks.
- Focused multimodal workflow: the agent uses rendered screenshots as part of an autonomous verify‑and‑iterate loop, illustrating practical patterns for vision‑guided code generation and self‑validation.
- Small, sanitized, MIT‑licensed dataset formatted for easy inspection with standard tools (pandas/polars), ideal as a concrete example rather than a large corpus.
Who It's For and Tradeoffs
Great fit if you study AI agent behavior, tool use, or multimodal self‑verification patterns — for researchers, educators, or engineers prototyping agent debugging workflows. Look elsewhere if you need large-scale labelled datasets, broad agent benchmarks, or raw production logs: this is a single, focused session (small size) and contains redactions/placeholders for private paths and environment details.
Where It Fits
Compliments larger agent‑trace collections by providing a clear, executable example of visual inspection integrated into an autonomous generation loop. Use it to prototype replay tools, visualization dashboards, or case studies about agent verification strategies.
