Doctorow's point hits hard when you look at how content and web interfaces are being churned out right now. The web is flooded with low-friction, high-noise garbage. As someone building a content generation tool, the biggest technical challenge isn't making the AI write more; it's making it write less. We explicitly instruct our models to be 'High Signal, Low Noise,' force short sentences, and ban vague platitudes. If a tool just generates walls of text to appease algorithms (or fill up modal dialogs), it's just contributing to the decay. Short, punchy, and grounded is the only way forward.
The most reliable heuristic humans use to detect AI isn't necessarily structural—it's vocabulary. We get a 'spidey sense' when we see certain words clustered together. I've spent the last few months tuning the generation engine for my text SaaS, and the single most effective way to make the output read as human isn't a complex prompt chain; it's a hard banlist. Forcing the model to never use words like 'Unlock', 'Delve', 'Supercharge', or 'In today's world' instantly changes the tone. Humans aren't really detecting AI; we're detecting RLHF-optimized lazy writing.
This is exactly why building AI tools for professional writing requires aggressive system-level guardrails. The default behavior of an LLM is to please the user, which often means fabricating a quote if it makes the narrative flow better. When building the generation engine for ConvertlyAI, we had to hardcode a strict rule into the system prompt: explicitly banning dummy testimonials, placeholder stats, or unsubstantiated claims unless they exist in the provided source context. If you don't force a strict 'Truth Check' constraint at the API level, the model will inevitably hallucinate a quote just to complete the 'story' structure it thinks you want.
We're currently using a strictly rule-based approach injected at the system-prompt level rather than a secondary classifier. Since ConvertlyAI handles 10 different output types (from raw Twitter threads to SEO blogs), we found that explicitly banning specific behaviors (like markdown wrappers, fake metrics, or specific generic AI buzzwords like 'Delve') directly in the main "systemRole" for "gpt" keeps latency low while still preventing formatting breaks. It's essentially a massive 'Do Not Do This' list passed right before execution.
Your static analysis approach for catching type constraints before execution sounds significantly more robust for complex workflows, though! Is that adding much latency?
Great read. I just pushed my own Show HN live about 10 minutes ago, so reading through your launch experience is perfectly timed anxiety-relief. Thanks for the transparency on your traffic spikes.
Congrats on shipping this. The subscription fatigue in the market right now is very real. It's exactly why I went with a hybrid pay-as-you-go credit model for my own SaaS instead of forcing a monthly sub on day one just to process an audio file. Did you build this natively for iOS, or are you using React Native/Expo to cross-platform it?
I love this approach to verification. I literally just launched my own AI formatting engine yesterday, and the hardest part wasn't the generation—it was building strict system-level guardrails to stop the model from outputting generic fluff words and breaking my slide formatting. Are you doing this pre-execution verification purely through secondary prompt checks, or are you running it through a separate smaller model first?