CrewAI Podcast episode planner & outline generator

This project uses CrewAI to simulate a podcast production team. AI agents collaborate to research a topic (and optional guest), structure an episode, generate talking points and questions, draft intro/outro snippets, and compile everything into a comprehensive episode plan.

This tool assists podcast creators by automating much of the pre-production planning process. It's designed to run in Google Colab.

Features

• Topic & guest research: Gathers background information, key facts, interesting angles, and guest expertise using web search.
• Episode structuring: Proposes a logical flow with distinct segments and estimated timings.
• Content generation: Creates relevant talking points and discussion prompts for each segment.
• Guest question drafting: Generates insightful, open-ended questions tailored to the guest's background (if applicable).
• Script snippet writing: Drafts host text for introductions and outros, including hooks and calls to action.
• Plan compilation: Assembles all generated content into a single, organized episode plan document (Markdown).

Requirements

• Python 3.9+
• Google Colab Environment (Recommended) or local Python setup.
• API Keys:
  • OpenAI API Key: For the language model (GPT-4 Turbo recommended for better quality). Requires a funded account or active credits from platform.openai.com.
  • Serper API Key: For the TopicResearcher agent's web search via SerperDevTool. Get one from serper.dev (free tier available).

Setup (Google Colab)

1. Get the Notebook: Clone the repository or download the .ipynb file.
2. Open in Colab: Upload and open in Google Colab (colab.research.google.com).
3. Install libraries: Run Cell 1 (# @title 1. Install...) to install necessary packages.
4. Configure API Keys in Colab Secrets:
   • Go to Colab Secrets (Key icon in the left sidebar).
   • Enable "Notebook access".
   • Add two secrets:
     • Name: OPENAI_API_KEY | Value: Your OpenAI key (sk-...)
     • Name: SERPER_API_KEY | Value: Your Serper key
   • Make sure the toggle switch is ON for both secrets.
5. Run Cell 2: Execute Cell 2 (# @title 2. Import Modules...). Check the output to verify both API keys were loaded successfully. Fix any errors before proceeding.

How to use

1. Define episode details (Cell 3):
   • Go to Cell 3 (# @title 3. Define Podcast Episode Details).
   • Modify the values for episode_topic, guest_name (set to None or "" if no guest), guest_info_url (if applicable), and podcast_style. Provide as much detail as relevant.
2. Select LLM (optional - Cell 4):
   • Ensure an appropriate model (GPT-4 Turbo recommended) is selected in Cell 4.
3. Run cells sequentially: Execute Cells 4 through 8 in order.
   • Cell 5 defines the agents.
   • Cell 6 defines the tasks.
   • Cell 7 creates the Crew and runs kickoff(). This is the main step and will take time and API calls. Observe the verbose output to see the agents work.
   • Cell 8 displays the final compiled episode plan.

Workflow overview

1. Topic researcher (Task 1): Researches the topic and guest (if any).
2. Episode structurer (Task 2): Creates the segment outline and timings.
3. Content point generator (Task 3): Develops talking points and guest questions for each segment.
4. Script element writer (Task 4): Drafts intro and outro text.
5. Plan compiler (Task 5): Assembles all parts into the final plan document.

Customization

• Episode details: Change inputs in Cell 3 for different episodes.
• Agent personalities: Modify role, goal, backstory in Cell 5.
• Task specificity: Adjust description, expected_output in Cell 6. For example, request more/fewer questions, different segment types, or a specific call to action.
• Structure/timing: Change the requirements in the EpisodeStructurer's goal/task description.
• Tools: Add or replace research tools if you have access to different APIs or databases.

Limitations

• API costs: LLM usage costs money. Monitor your OpenAI spending.
• Research quality: Depends on the search tool's effectiveness and the LLM's ability to synthesize information accurately. Web scraping/reading can be unreliable.
• Creativity: While capable, the AI's creativity for hooks, questions, or unique angles might be limited compared to a human producer.
• Guest nuance: Understanding a guest's true personality or subtle expertise from online info alone is difficult.
• This is an assistant: The generated plan is a strong starting point but should be reviewed, edited, and adapted by the actual podcast host/producer.