I built ConvertlyAI (https://convertlyai.online) to solve my own post-production bottleneck. As a writer and creator, I found that producing the core content—whether a long-form manuscript or a video—was only half the battle. Chopping that single piece of content up into highly specific formats for Twitter threads, LinkedIn carousels, and newsletters was causing massive burnout.

I wanted a system where I could upload a raw YouTube transcript or an MP3, and have an engine strictly format it for 10 different platforms without losing the core message or sounding like a robot.

How it works under the hood: The stack is Next.js, Supabase, and the OpenAI SDK (GPT-4o).

The biggest challenge wasn't generating text; it was preventing the AI from generating garbage text. I spent a lot of time engineering strict system-level guardrails to force the model to be "High Signal, Low Noise."

Banned Vocabulary: I hardcoded rules to strip out generic AI fluff words (e.g., "Unlock," "Delve," "In today's fast-paced world," "Supercharge").

Strict Formatting: Each of the 10 outputs has a dedicated sub-prompt. For example, the Twitter thread logic forces "1/X" numbering and bans hashtags in the hook, while the LinkedIn logic strictly formats text for slide carousels.

Audio Processing: I built a transcription pipeline so users can just upload raw MP3s if they don't have text ready.

I'm currently dogfooding the product to automate my own content distribution and running a credit-based billing system to manage the compute costs for the AI and audio processing.

I would love for you to tear it apart. I'm specifically looking for feedback on:

The UI/UX of the dashboard when managing saved "Magic Prompts."

The strictness of the output formatting—if you run a transcript through it, does it actually sound human to you?

Happy to answer any questions about the prompt engineering, the tech stack, or the API routes!