Open source · LLM use case

GhostWhisper

A Telegram bot that writes blog posts for you: describe an idea by text or voice message and it publishes a draft article to your Ghost blog. I built it in a weekend as a hands-on use case to explore prompt management for LLM apps, and turned it into a talk.

Source on GitHubSlides (PDF)

How it works

1. You send the bot a short description of the article, as text or as an audio message.
2. Audio is transcribed with OpenAI Whisper, in any language.
3. ChatGPT turns the description into an article, using prompts managed in Pezzo (for example a casual Spanish or Polish style).
4. The article is created in Ghost through its Admin API.

Stack

• Bun / Node.js with Telegraf for the Telegram bot.
• Ghost as the blogging platform.
• Whisper and ChatGPT from OpenAI.
• Pezzo for prompt versioning, environments and cost monitoring.
• Docker Compose to run the bot and a Ghost instance together.

The idea

Let the LLM implement the logic and focus on designing and tuning the prompts. Keeping prompts outside the code, with versions, tests and separate environments, makes them as manageable as any other part of the application.

Run it

Copy .env.example to .env with your Telegram, OpenAI, Pezzo and Ghost keys, then start everything with docker-compose up -d. The repository explains how to get each key.

This is a prototype from 2023, built to learn and to share, not a maintained product.