Claude Code /init on an Existing Project

Today I pointed Claude Code at a small existing project — my Streamlit and Gradio demo apps that run on apps.databloom.net. I started Claude Code, ran /init, and let it read through the files. It produced the CLAUDE.md below entirely on its own — I didn't give it any other instructions.

Everything it captured — the project overview, the development commands, the architecture, the key patterns — it got all of that from reading the source files. Not bad for a first pass on a codebase it had never seen before.

The Generated CLAUDE.md

# CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

## Project Overview

DataBloom.net demo apps — a portfolio of AI/GenAI and Python demos by Jeremy Bloom. Two web apps
(Streamlit and Gradio) run on a single EC2 instance behind Nginx, managed by systemd.

- **Streamlit app**: https://apps.databloom.net
- **Gradio app**: https://gradio.apps.databloom.net

## Development Commands

**Prerequisites:** Python 3.12 and uv (https://github.com/astral-sh/uv). Python 3.13 is not
supported (NumPy compatibility).

### Streamlit
    cd streamlit
    uv venv -p 3.12 && source .venv/bin/activate
    uv pip install -r requirements.txt
    streamlit run app.py                    # runs on http://localhost:8501

### Gradio
    cd gradio
    uv venv -p 3.12 && source .venv/bin/activate
    uv pip install -r requirements.txt
    python app.py                           # runs on http://localhost:7860

### Tests
    cd streamlit
    pytest pages/test_benchmark_problems.py

### Deployment (on EC2)
    sudo systemctl restart streamlit
    sudo systemctl restart gradio-demo

## Architecture

### Streamlit App (streamlit/)
Multi-page Streamlit app. app.py is the entry point; pages live in pages/ with numeric prefixes
controlling sidebar order (001–012). Pages range from utility demos (logger, timer) to AI-powered
features (chat, debate, benchmarking).

**Core modules:**
- chatting.py — Unified LLM interface. chat_with_llm() routes to provider-specific functions
  (chat_with_gpt(), chat_with_anthropic()) based on model name prefix.
- support.py — Model enumeration; defines available models per provider.
- rate_limiter.py — Sliding-window rate limiter using deque. Stored in st.session_state.
- response_parser.py — Whitelist-based field extraction from API responses (prevents leaking
  sensitive data). Includes token usage and cost estimation.
- logger.py — Session-aware logging with UUID-based session IDs. Writes to app.log.

**LLM providers supported:** OpenAI, Anthropic, plus alternatives via environment keys (Gemini,
DeepSeek, Groq, Grok, OpenRouter).

### Gradio App (gradio/)
Minimal starter app — a single app.py with a hello-world Interface.

## Key Patterns

- **API keys** are loaded from streamlit/.env via python-dotenv. Keys exist for OpenAI, Anthropic,
  Gemini, DeepSeek, Groq, Grok, and OpenRouter.
- **Session state** is managed through st.session_state for rate limiters, conversation history,
  and session IDs.
- **Dependencies** are managed with uv. requirements.txt is the source of truth; pyproject.toml
  and uv.lock also exist per app.
- Each Streamlit page links back to its GitHub source for educational purposes.
- benchmark_problems.json contains structured logic puzzles (with difficulty levels and categories)
  used by 011_ai_benchmark.py for LLM evaluation.

← Back to Home