Prompt Engineering Workbench | Claude Projects

A powerful workbench for testing and refining Claude-powered workflows, agents, and prompts.

⚠️ SECURITY WARNING: This tool is designed for local development only. Do NOT expose it to your network or the internet without adding proper authentication. It has no built-in access controls and anyone who can reach the server can use your API credentials and access your conversations.

Why This Tool?

Claude Projects lets you create custom AI assistants with specific instructions and knowledge bases. But testing those workflows requires switching between the Claude.ai interface, copying prompts, and manually iterating.

This workbench solves that by giving you:

⚠️ Note: This tool uses cookie-based authentication to access Claude.ai directly because the Claude Projects API is currently in beta and not publicly available. Once Anthropic releases the official Projects API, this tool can be updated to use proper API authentication instead of browser cookies. Until then, this is a practical workaround for teams who need to test their Projects programmatically.

Direct access to your Claude Projects - Test your agents with their full knowledge bases
Reusable prompt templates - Build, save, and iterate on prompts with variables
Tool execution - Claude can fetch web pages and search the web during analysis
Progress visibility - See what Claude is doing in real-time

Use Cases

Mode	Best For
Projects	Testing customer-facing agents, workflows with knowledge bases, complex multi-step prompts
API	Simple prompt templating, quick iterations, no knowledge base needed

Features

Core Features

Dual Connection Modes - Use official API or direct Claude.ai web access
Mode Selector - Switch between API and Projects mode in Settings (when both configured)
Claude Projects Support - Access your Claude Projects with their knowledge bases
Project Selector - Switch between multiple Claude Projects on the fly
Configuration-driven - Works with any Claude Project by editing YAML config
20+ Built-in Templates - Analysis, research, content creation, and more

Tool Execution

Web Fetch - Claude can fetch and analyze web pages automatically
Web Search - DuckDuckGo (default) or Google Custom Search integration
Progress Tracking - Visual progress bar with tool execution stats
Real-time Feedback - See fetches and searches as they happen

File & Content

File upload support - PDF, DOCX, TXT, MD (configurable)
Drag & drop uploads - Easy file handling
URL content fetching - Optionally fetch and analyze web content

UI/UX

Template System - Dynamic prompt builder with variable inputs
Conversation history - Maintains context within sessions
New Chat button - Start fresh conversations within projects
Settings modal - Manage projects, templates, and connection settings
Progress indicators - Animated progress bar during processing
Responsive design - Works on desktop and mobile

Connection Modes

Projects Mode (Recommended for Agent Testing)

Connect directly to Claude.ai using your browser session cookie. This is the primary use case for testing customer workflows and agents.

Why use Projects mode:

Test agents with their full knowledge bases and custom instructions
Validate complex workflows before deploying
Iterate on project configurations with real feedback
Access project-specific conversations and context

Header displays: Prompt Engineering Workbench | Claude Projects

API Mode (Simple Prompt Templating)

Use the standard Anthropic API for direct Claude access. Best for users who just need prompt templating without project features.

Why use API mode:

Simple prompt engineering and iteration
No knowledge base or custom instructions needed
Direct API access for testing raw prompts
Lower friction for quick experiments

Header displays: Prompt Engineering Workbench | Claude AI

Switching Modes

If you have both an API key and a Claude.ai cookie configured, you can switch between modes:

Click the Settings (gear icon) in the header
In the Connection section, you'll see mode toggle buttons
Click Projects or API to switch modes
The UI will update automatically (header title, project selector visibility)

Quick Start

1. Clone and Navigate

git clone https://github.com/SEMalytics/claude_project_chat.git
cd claude_project_chat

2. Run Setup (Recommended)

The easiest way to get started:

./setup.sh

This script will:

Check Python version (requires 3.9+)
Create a fresh virtual environment
Install all dependencies
Create .env from .env.example if needed

3. Manual Setup (Alternative)

If you prefer to set up manually:

# Remove any existing venv (if reinstalling)
rm -rf venv

# Create fresh virtual environment
python3 -m venv venv

# Install dependencies using venv's pip directly
./venv/bin/pip install --upgrade pip
./venv/bin/pip install -r requirements.txt

Important: Always use ./venv/bin/pip and ./venv/bin/python to ensure you're using the virtual environment's Python, not your system Python.

4. Configure Environment

# Copy example env file
cp .env.example .env

# Edit .env with your credentials
nano .env  # or: code .env, vim .env, etc.

For Claude.ai Web Client (Recommended):

# Get this from your browser's DevTools (see instructions below)
CLAUDE_COOKIE=sessionKey=sk-ant-sid01-xxx...

# Optional: Start in a specific conversation
CLAUDE_CONVERSATION_ID=your-conversation-uuid

Getting Your Claude Cookie:

Open claude.ai and log in
Open DevTools (F12 or Cmd+Option+I)
Go to Network tab
Click any request to claude.ai
Find the Cookie header in Request Headers
Copy the entire cookie string (starts with sessionKey=)

For Official Anthropic API:

ANTHROPIC_API_KEY=sk-ant-api03-your-key-here
CLAUDE_PROJECT_ID=your-project-uuid-here

Optional: Google Custom Search (for web search tool)

By default, web search uses DuckDuckGo (no API key needed). For Google Custom Search:

# Get API key from: https://console.cloud.google.com/apis/credentials
GOOGLE_SEARCH_API_KEY=your-google-api-key

# Create search engine at: https://programmablesearchengine.google.com/
GOOGLE_SEARCH_CX=your-search-engine-id

5. Run the Server

# Using venv python directly (recommended)
./venv/bin/python app.py

# OR activate venv first, then run
source venv/bin/activate  # macOS/Linux
python app.py

You should see:

==================================================
  Prompt Engineering Workbench | Claude Projects

==================================================

  Server starting at http://127.0.0.1:5001
  Press Ctrl+C to stop

6. Open in Browser

Navigate to: http://127.0.0.1:5001

7. Enable Projects (First Time)

Click the Settings (gear icon) in the header
In "Active Projects", check the projects you want to use
Click Save Changes
Use the project dropdown in the header to switch projects

Configuration

Project Settings (`project_config.yaml`)

Edit this file to customize the interface:

# Project Identity
project:
  name: "My Project"
  description: "Description here"

# UI Customization
ui:
  title: "Chat Title"
  subtitle: "Subtitle text"
  primary_color: "#3b82f6"

# Features
features:
  file_upload: true
  url_fetching: true
  multi_file: true

# File Settings
files:
  allowed_extensions: [pdf, docx, txt, md]
  max_size_mb: 10
  max_files: 5

# Custom Prompts (for Simple Input mode)
prompts:
  - id: "general_chat"
    label: "General Chat"
    template: "{user_input}"
    requires_files: false
    placeholder: "Ask anything..."

Custom Templates (`custom_prompts.yaml`)

Create your own templates that persist across updates:

templates:
  - id: my_custom_template
    name: "My Custom Analysis"
    description: "Describe what this does"
    category: analysis
    template: |
      Analyze the following for [focus_area]:

      [content]
    variables:
      - name: focus_area
        label: "Focus Area"
        type: text
        required: true
      - name: content
        label: "Content to Analyze"
        type: textarea
        required: true

Project Structure

claude_project_chat/
├── app.py                   # Flask server with API routes
├── config.py                # Configuration loader
├── project_config.yaml      # Project-specific settings
├── custom_prompts.yaml      # Your custom templates (gitignored)
├── requirements.txt         # Python dependencies
├── .env                     # API keys/cookies (gitignored)
├── .env.example             # Example environment file
├── static/
│   ├── css/styles.css       # Custom styles
│   ├── js/
│   │   ├── app.js           # Main frontend JavaScript
│   │   └── prompt-builder.js # Template UI + Project Manager
│   └── uploads/             # Temporary file storage
├── templates/
│   └── index.html           # Chat interface template
└── utils/
    ├── __init__.py
    ├── claude_client.py     # Official Anthropic API wrapper
    ├── claude_web_client.py # Claude.ai web client (cookie auth)
    ├── tool_handler.py      # Tool execution handler
    ├── file_processor.py    # File handling
    ├── url_fetcher.py       # URL content fetching
    ├── prompt_templates.py  # 20+ built-in templates
    └── prompt_compiler.py   # Template compilation

API Endpoints

Endpoint	Method	Description
`/`	GET	Main chat interface
`/api/chat`	POST	Send message to Claude
`/api/upload`	POST	Upload file
`/api/fetch-url`	POST	Fetch URL content
`/api/templates`	GET	Get all templates
`/api/prompts`	GET	Get legacy prompts
`/api/session/{id}`	GET	Get session history
`/api/session/{id}`	DELETE	Clear session
`/api/config`	GET	Get project config
`/api/client-status`	GET	Check connection mode and available options
`/api/switch-mode`	POST	Switch between API and Web modes
`/api/update-cookie`	POST	Update Claude.ai cookie
`/api/projects`	GET	List available projects
`/api/projects/set-active`	POST	Switch active project
`/api/conversations`	GET	List conversations
`/api/conversations/new`	POST	Start new conversation

Built-in Template Categories

Analysis - SWOT, competitive, gap analysis
Research - Literature review, market research, trend analysis
Content - Blog posts, social media, email campaigns
Technical - Code review, documentation, architecture
Business - Business plans, proposals, reports
Creative - Brainstorming, storytelling, naming

Troubleshooting

Cookie Authentication Issues

"Access denied (403)"

Your cookie has expired. Get a fresh one from claude.ai DevTools
Update via Settings modal or edit .env directly

Projects not showing

Click Settings (gear icon)
Click "Refresh Projects"
Enable the projects you want
Save Changes

API Issues

"ANTHROPIC_API_KEY is required"

Check that .env file exists and contains your API key
Ensure the key starts with sk-ant-

File Upload Issues

Check file extension is in allowed_extensions
Verify file size is under max_size_mb
Ensure static/uploads/ directory exists

Module Not Found

# Activate virtual environment
source venv/bin/activate

# Reinstall dependencies
pip install -r requirements.txt

Pip Installation Issues

If pip installs to wrong location:

# Use venv binaries directly
./venv/bin/pip install -r requirements.txt
./venv/bin/python app.py

Development

Format Code

pip install black flake8
black .
flake8 .

Run Tests

pip install pytest
pytest

Security Notes

API keys and cookies are stored in .env (not committed to git)
Runs locally on 127.0.0.1 by default
Uploaded files are stored temporarily in static/uploads/
Cookie-based auth uses your Claude.ai session - keep it secure
Consider adding authentication for production use

Credits & Acknowledgments

This project uses and was inspired by several open-source projects:

Libraries Used

Flask - Web framework
Anthropic Python SDK - Official Claude API
curl_cffi - HTTP client with browser impersonation
duckduckgo-search - Web search without API key
BeautifulSoup4 - HTML parsing
PyPDF2 - PDF text extraction
python-docx - Word document processing
Tailwind CSS - Utility-first CSS framework

Inspiration

Claude-API by KoushikNavuluri - Inspired the approach for direct Claude.ai web access using cookie authentication

Built With

Claude Code - AI-assisted development

License

MIT License - Use freely for your projects.

Note: This is an unofficial tool and is not affiliated with Anthropic. Use responsibly and in accordance with Anthropic's terms of service.

About SEMalytics

This tool is built by SEMalytics—making AI development workflows practical.

Related Projects:

Forge — AI code generation with 12-agent review system
KnowledgeForge — Multi-agent orchestration methodology
Concept Clarity — Message clarity measurement system

Learn more at SEMalytics →

_{Built by SEMalytics · AI tools that ship.}

Name		Name	Last commit message	Last commit date
Latest commit History 27 Commits
static		static
templates		templates
utils		utils
.env.example		.env.example
.gitignore		.gitignore
CHAT_INTERFACE_PLAN.md		CHAT_INTERFACE_PLAN.md
CLAUDE.md		CLAUDE.md
README.md		README.md
app.py		app.py
config.py		config.py
project_config.yaml		project_config.yaml
requirements.txt		requirements.txt
setup.sh		setup.sh

SEMalytics/claude_project_chat

Folders and files

Latest commit

History

Repository files navigation