Noren

Today we published the blog post we've been building toward: how we spent weeks writing a 300-line voice guide by hand, fed it to an LLM, and it finally sounded right. Then why we built an engine to do it automatically.

We also formally introduced Noren on X today. Voice extraction for AI writing.                   Launching April 6. Waitlist open at usenoren.ai. 

Earlier in the week I posted about how our first blog post got flagged for AI-ism patterns. Triple parallel structures, hedge qualifiers, copula dodging. The patterns that make readers think "this is AI" even when the content is good. (https://www.indiehackers.com/post/our-first-blog-post-got-called-out-for-ai-isms-touche-865133f220).

So we built detection for it directly into Noren's output checks. Four detectors scanning every generation: triple parallels, gerund litanies, hedge qualifiers, and copula dodging. Pure string matching, no LLM calls.

Violations get flagged as metadata alongside our existing anti-pattern checks. Building in public means your users call you out and you ship the fix the same week.

Why all AI writing sounds the same (and what's actually causing it)

We just published our first blog post breaking down the real reason AI output sounds generic. It's not a prompt engineering problem. It's a pattern problem.

LLMs default to the same sentence structures, vocabulary, and rhythms regardless of what you tell them in the system prompt. "Write in a professional but friendly tone" produces identical output for every user.

Full post: https://usenoren.ai/blog/why-ai-writing-sounds-the-same 

Curious if this matches what others here have experienced.

We built Noren to solve a problem that was driving us crazy: the “AI Editing Tax.”

Like most Indie Hackers, we started using LLMs to draft emails, tweets, and investor updates. The generation took 30 seconds. But the output was... generic. It sounded like a hyper-enthusiastic ghostwriter. We'd then spend 15 minutes rewriting the draft so it sounded like us—killing the semicolons, changing the vocabulary, fixing the sentence rhythms.

The promise was speed, but we were just trading writing time for editing time.

To fix this, I tried doing it the hard way. I spent three weeks analyzing my own writing samples and documenting every single pattern. I ended up with a 300-line Markdown voice guide covering my analogy domains, rhetorical moves, and punctuation habits. When I fed that to Claude, the output finally clicked. It sounded like me.

But looking at that 300-line file, I realized: this is just pattern recognition.

If I can manually extract these patterns over three weeks, an engine should be able to do it automatically. So we built that engine.