TOGEN.md

Final Plan: The "TOGEN" AI Podcast Prompt Engineer

The application is a command-line tool that intelligently generates a customized, high-quality prompt file. This prompt is designed to be used in a separate LLM (like NotebookLM) to generate a podcast script.

---

Phase 1: Project Setup & CLI Definition

1. Cargo.toml Dependencies:

   • tokio: For the asynchronous runtime.
   • serde & serde_json: For robust JSON serialization and deserialization of our component files.
   • clap: For creating a clean and powerful command-line interface.
   • reqwest: A reliable HTTP client for communicating with the local LLM API.
   • thiserror & `anyhow``: For ergonomic and easy-to-manage error handling.

2. CLI Structure (src/main.rs):

   • The CLI will accept the following arguments:
     • file_path: Mandatory path to the source content file (e.g., raw/news/some-ai-news.html).
     • --output-dir: An optional path to a directory for saving the generated prompt. Defaults to a new ./generated_prompts directory.
     • --llm-url: An optional URL for the local LLM endpoint, defaulting to http://localhost:9090/prompt.

---

Phase 2: Core Logic & Execution Flow

The tool will follow a "Reason -> Assemble -> Save" workflow.

1. Read Source File: Read the full content of the file specified by file_path.

2. Intelligent Format Selection (Single LLM Call):

   • The tool sends a lightweight "meta" prompt to the local LLM.
   • Prompt: "Analyze the following content and determine the best podcast format: 'news_summary', 'explainer', or 'paper_deep_dive'. Respond with only the chosen format name."
   • This determines which blueprint from the prompt_components/formats/ directory to use.

3. Load All Prompt Components:

   • Based on the format chosen by the LLM, the application reads and deserializes all the required JSON files (core, personas, segments, rules, and the selected format) into Rust structs.

4. Final Prompt Assembly:

   • A builder function assembles the final prompt string. The structure will be podcast_template + instructions + file_content to properly prime the target LLM.
   • The generated Markdown will be clean and token-efficient, using simple headers (e.g., # Title, ## Subtitle).
   • The final prompt will look like this:

     # CORE IDENTITY
     Podcast Name: Noob Learning
     Slogan: Noob Learning: Let's vibe learning together!
     ...
     
     # HOST PERSONAS
     ...
     
     # GOLDEN RULES
     ...
     
     # EPISODE BLUEPRINT: AI News Summary
     Your goal is to summarize a recent AI news story...
     ## The News Hook (30s)
     ...
     
     ---
     # SOURCE CONTENT
     [The full content from the input file is placed here]

5. Save the Final Prompt:

   • The assembled prompt is saved to a new .md file.
   • The filename will be descriptive, combining the source filename, the chosen format, and the current date, e.g., some-ai-news-news_summary-prompt-2023-10-27.md.
   • The file will be saved in the specified output directory.
   • The application will print a confirmation message to the console indicating where the file was saved.