Claude API Features — Jupyter Notebook Demos

A hands-on demonstration suite for key Claude API capabilities, implemented as Jupyter notebooks. Each notebook is self-contained and focuses on a single feature.

Features Covered

Notebook	Feature	Description
001_thinking_complete.ipynb	Extended Thinking	Surface Claude's internal reasoning before a final answer
002_citations.ipynb	Citations	Get source-attributed answers grounded in a document
002_images.ipynb	Vision / Image Analysis	Analyze satellite images for structured risk assessment
002_pdfsupport.ipynb	PDF Processing	Parse and summarize PDF documents directly
003_caching.ipynb	Prompt Caching + Tool Use	Cache large system prompts and tool definitions to cut costs
005_code_execution.ipynb	Code Execution + Files API	Execute Python in a sandbox, upload/download files

---

Prerequisites

• Python 3.9+
• Jupyter Lab or Jupyter Notebook
• An Anthropic API key

---

Setup

1. Clone the repository

git clone https://github.com/<your-username>/ClaudeFeatures.git
cd ClaudeFeatures

2. Create and activate a virtual environment

python -m venv .venv

# Windows
.venv\Scripts\activate

# macOS / Linux
source .venv/bin/activate

3. Install dependencies

pip install anthropic python-dotenv jupyter

4. Configure your API key

Copy the .env file template and add your key:

# .env
ANTHROPIC_API_KEY=your_api_key_here

Note: .env is listed in .gitignore and will never be committed.

5. Launch Jupyter

jupyter lab

---

Notebook Details

001_thinking_complete.ipynb — Extended Thinking

Enables Claude's internal reasoning (ThinkingBlock) before it produces a final TextBlock answer. Useful for complex problems where showing work matters.

Key API parameters:

thinking={"type": "enabled", "budget_tokens": 10000}

What to look for: The response object contains both a ThinkingBlock (redacted reasoning trace) and a TextBlock (final answer).

---

002_citations.ipynb — Citations

Passes an Earth article as a document and asks a factual question. Claude returns its answer with inline citations — each one carries the exact quoted text, document title, and character offsets.

Key API parameters:

{"type": "document", "source": {...}, "title": "...", "citations": {"enabled": True}}

What to look for: CitationCharLocation objects in the response with cited_text, start_char_index, and end_char_index.

---

002_images.ipynb — Vision / Image Analysis

Sends a base64-encoded satellite image (images/prop7.png) to Claude with a structured fire risk assessment prompt. Claude identifies the residence, evaluates tree overhang and defensible space, and returns a 1–4 risk rating.

Key API parameters:

{"type": "image", "source": {"type": "base64", "media_type": "image/png", "data": "<base64>"}}

What to look for: Structured property analysis in plain text with an explicit risk score.

---

002_pdfsupport.ipynb — PDF Processing

Encodes earth.pdf as base64 and asks Claude to summarize it. Demonstrates that Claude can directly parse PDF content without a separate extraction step.

Key API parameters:

{"type": "document", "source": {"type": "base64", "media_type": "application/pdf", "data": "<base64>"}}

What to look for: A one-sentence summary produced from the raw PDF bytes.

---

003_caching.ipynb — Prompt Caching + Tool Use

Wraps a large JavaScript-generation system prompt (~6k tokens) and four tool definitions with cache_control so they are reused across turns. Reduces cost significantly for multi-turn sessions with large, static context.

Key API parameters:

# On the last tool definition and on the system prompt block:
"cache_control": {"type": "ephemeral"}

What to look for: usage.cache_creation_input_tokens on the first call, then usage.cache_read_input_tokens on subsequent calls — the second request costs far fewer tokens.

Tools defined:

• add_duration_to_datetime — date/time arithmetic
• set_reminder — schedule a notification
• get_current_datetime — current timestamp
• db_query — SQLite query execution

---

005_code_execution.ipynb — Code Execution + Files API

Uploads streaming.csv (a customer churn dataset) via the Files API, asks Claude to analyze it and generate a visualization, then downloads the resulting chart and report.

Beta headers required:

"anthropic-beta": "code-execution-2025-08-25, files-api-2025-04-14"

Workflow:

1. Upload CSV → get a file_id
2. Pass file_id in the message with a code_execution_20250825 tool
3. Claude writes and runs Python in a sandboxed environment
4. Download the output chart (churn_analysis_dashboard.png) and report via file_id

What to look for: The notebook demonstrates the full lifecycle — upload, execute, retrieve — and shows how each code execution block starts with a clean state (no persistent variables between blocks).

---

Project Structure

ClaudeFeatures/
├── .env                          # API key (not committed)
├── .gitignore
├── LICENSE                       # MIT
├── README.md
├── 001_thinking_complete.ipynb
├── 002_citations.ipynb
├── 002_images.ipynb
├── 002_pdfsupport.ipynb
├── 003_caching.ipynb
├── 005_code_execution.ipynb
├── earth.pdf                     # Sample document (citations + PDF demos)
├── streaming.csv                 # Sample churn dataset (code execution demo)
├── churn_analysis_dashboard.png  # Output chart from code execution demo
└── images/
    ├── prop1.png
    ├── ...
    └── prop7.png                 # Satellite images (vision demo)

---

Common Patterns

All notebooks follow the same structure:

from dotenv import load_dotenv
import anthropic, os

load_dotenv()
client = anthropic.Anthropic(api_key=os.getenv("ANTHROPIC_API_KEY"))

messages = []

def add_user_message(content):
    messages.append({"role": "user", "content": content})

def chat(**kwargs):
    response = client.messages.create(
        model="claude-sonnet-4-5",
        max_tokens=8096,
        messages=messages,
        **kwargs
    )
    messages.append({"role": "assistant", "content": response.content})
    return response

Feature-specific parameters (thinking, tools, caching headers) are passed as keyword arguments to chat().

---

License

MIT © 2026 Narmada Nannaka