Explore PaperBanana codebase to learn figure generation workflow
Session from traceweave on Hugging Face by Samarth0710 · MIT · condensed by Coders Talk
Click a segment to jump to that moment
1msession
0interventions
GOAL
Understand the PaperBanana repo — what it is, how to generate figures, its available scripts/tools, and its documentation — by reading through the code and docs.
- +0Prompt · +0Explore the paperbanana directory thoroughly: what it is, how to generate figures, available tools/scripts, docs, and supported figure types.
- +0mAgent did · +0mSearched for docs/config files and read README.md, docs/paperbanana.md, requirements.txt, and directory listing to get an overview.
- +0mAgent did · +0mRead paperbanana.py, config.py, and all agent modules (retriever, planner, visualizer, stylist) to understand the multi-agent pipeline architecture.
- +0mAgent did · +0mRead critic.py, aesthetic_guidelines.py, and examples.py to understand the refinement loop and styling guidelines.
- +0mAgent did · +0mInspected load_reference_set.py, generate_paper_figures.py, and app.py (the Gradio HuggingFace Spaces interface) to see end-to-end usage.
- +0mAgent did · +0mExplored agents/__init__.py, data/ and examples/ directories, utils.py, scripts/ folder, and enumerated all Python files in the repo.
- +0mAgent did · +0mRead scripts/generate_self_diagram.py and sampled the spotlight_reference_set.json to see the reference example format.
- +1mOutcome · +1mProduced a comprehensive written overview of PaperBanana: its 5-agent pipeline (Retriever, Planner, Stylist, Visualizer, Critic), usage scripts, and supported figure/diagram types.
VERDICT
I'd start with a more targeted prompt pointing straight at the entry-point scripts (paperbanana.py, app.py, generate_paper_figures.py) instead of asking for a full directory sweep, since most of the value came from the agent files anyway. Letting it read everything did work here with no corrections needed, so for a pure exploration task like this the thoroughness paid off.