I built a tool that filters AI slop out of English social posts. The hardest part was teaching AI to stop sounding like AI.

Fair, one ugly before/after beats five polished claims.

Before: "Our platform leverages cutting-edge AI to seamlessly empower your workflow."
After: "Paste your draft, it flags the AI tells and shows why each one reads fake."

The em dash is the easy catch. Harder one was Chinese: AI writes 深入探讨, humans write 聊聊X. Not a word swap, AI picks the formal register even when the context is casual.

Moving one concrete transformation above the fold next. Curious how PostPilot handles the "invisible logic" thing, do you show why each change happened?

I’d trust this faster if you lead with the exact before/after on the em dash cleanup plus one Chinese phrase you adapted badly at first, that part felt real. I built PostPilot after running into the same textbook ChatGPT tone problem, and the stuff people trust quickest is usually one ugly raw input, one cleaned output, and one note on why the rewrite changed. Free isnt the issue imo, invisible logic is, can visitors see one concrete transformation above the fold?

Rhythm over authenticity is the better frame. I was scoring for authenticity and getting false positives on informal writing. Switched to rhythm metrics (sentence length variance, clause structure variance) and the signal got cleaner. Your jagged vs even description is exactly right.

On dual-model scoring: I tried Claude as judge and GPT as judge on the same outputs. The disagreements are where the gold is. Claude catches hedging patterns ("it seems", "arguably") more aggressively. GPT flags structural repetition better. Where they both agree, those are the reliable tells. Where they disagree, those edge cases forced me to write regex rules I would have missed otherwise.

Biggest surprise: GPT-4o gives higher scores to its own output than Claude does. Not shocking but the consistency of the bias was interesting. About 15% of the time GPT rates its own slop as 90+ while Claude catches it at 60. That delta is basically a map of what each model is blind to.

Rhythm over authenticity is the better frame. I was scoring for authenticity and getting false positives on informal writing. Switched to rhythm metrics (sentence length variance, clause structure variance) and the signal got cleaner. Your jagged vs even description is exactly right.

On dual-model scoring: I tried Claude as judge and GPT as judge on the same outputs. The disagreements are where the gold is. Claude catches hedging patterns ("it seems", "arguably") more aggressively. GPT flags structural repetition better. Where they both agree, those are the reliable tells. Where they disagree, those edge cases forced me to write regex rules I would have missed otherwise.

Biggest surprise: GPT-4o gives higher scores to its own output than Claude does. Not shocking but the consistency of the bias was interesting. About 15% of the time GPT rates its own slop as 90+ while Claude catches it at 60. That delta is basically a map of what each model is blind to.

the naming feedback is thoughtful. you are right that VoiceBridge sounds like a voice or translation product to someone who does not know the context. the trust layer vs writing assistant framing is interesting too, though i think for now the product is still closer to a utility than a platform people trust with their identity. Beryxa is clean and pronounceable but i worry about abstract names in a category where nobody knows the category exists yet. descriptive names like Notion or Linear work because the product category is already established. for something new, i might need the name to hint at what it does until people know what to search for. the bar you set, can a non-native founder say it, remember it, trust it, is the right one. still working on meeting it.

The 'teach AI to stop sounding like AI' framing is the right end of the problem. Spent a week last month forcing Claude to drop em-dashes, the rule-of-three, and the soft 'this isn't just X, it's Y' construction. About 60 percent of the AI-ness survives because the model's pretraining is the signal. The other 40 percent comes out in the rewrite. What was your hardest pattern to suppress?

The three-layer approach is exactly right. We ran into the same problem.

The regex cleanup layer ended up being the most reliable for us too. One thing I'd add: the AI-tell list ages out faster than you expect. New tells emerge quarterly as readers start recognizing current patterns and models adapt. Keeping that list live is ongoing maintenance work, not a one-time build.

The 'coaching ending' problem also runs deeper than just the last sentence. There is a mid-post version of it: 'X is one thing. But here is what really matters.' or 'The real question is not X, it is Y.' Those middle-of-post pivots are structural, not a fixed phrase, so they are harder to catch with regex. LLM-scored detection handles those better than pattern matching.

On marketing: free tools without signup face a 'why does this exist' trust gap. The Chinese internet slang table (neijuan, chuhai, etc.) is your strongest concrete proof that you solved an actual problem, not just called an API. Putting two or three of those examples on the landing page above the fold would do more than any feature list.

The AI-ness scoring concept is genuinely clever. Most people building AI writing tools are trying to make AI sound better at being AI. You're doing the opposite, making it sound less like AI, and that's actually the harder technical problem.

For marketing: your biggest asset is your own story. A Chinese developer who couldn't post naturally in English, so he built a tool to solve it. That's the kind of thing people share. Post before/after comparisons of the same thought through ChatGPT vs VoiceBridge on X. The visual contrast IS the marketing.

(I'm Aisa from aisa.to, we do AI skills assessment, so the quality-of-AI-output angle is something we think about a lot from the assessment side.)

The trust barrier for free tools is rarely about price. It's about uncertainty. With a writing tool, the specific fear is: 'does this store my drafts and sell training data?' Being explicit about 'no account, nothing stored' has to be in the first sentence, not buried in a FAQ. If I have to look for that information, I've already hesitated.

The platform-specific voice adaptation is the real differentiator here. Every other AI writing tool treats LinkedIn and X as the same problem. They're not. What you've built deserves that at the center of the positioning, not 'clean AI score.'

For distribution, the audience you want isn't indie hackers who write in English. It's ESL tech founders and builders posting in competitive markets (Indian, Chinese, Brazilian indie communities on X). They have the exact problem you described and they know it. A few native posts from real users in those communities will do more than any Product Hunt launch.

We ran into the same distribution wall at 3vo.ai when launching a tool with no signup. What finally moved the needle: seeding with 5-10 real before/after examples from actual users, not made-up demos. People trust proof of someone else's experience more than a badge that says 95/100.

What's the split between people who land on the tool and paste something vs. bounce immediately?

The “cleanup only” point is where I’d draw the product boundary. If the tool promises to make people sound native, users will expect it to rewrite their thinking. If it promises to preserve the thought and remove platform friction, the trust bar is much lower.

For the landing page, I’d show before/after examples by platform immediately above the CTA. Same raw Chinese note, then X / LinkedIn / Reddit outputs side by side. The score is interesting, but the proof is whether I can recognize the original thought after the rewrite.

The hardest part you're describing - teaching AI to not sound like AI - is the exact problem that makes AI content invisible in a different direction: AI readers (ChatGPT, Perplexity, Claude) are increasingly pattern-matching 'AI slop' and deprioritizing it too.

We started noticing this with 3vo.ai's content. Posts that were too polished, too structured in the 'AI way' were getting lower pickup in AI-generated answers about our category - even when the facts were accurate.

The irony: slop-filtered content (that sounds genuinely human) probably performs better in AI visibility too, not just human readability. The signal that makes it readable to humans - specificity, original observations, non-generic framing - is also what gets it cited by AI.

So you're solving two problems at once. The use case might be bigger than social posts.

That positioning advice is spot on. "Platform-native English posts" is exactly what I'm going for but I couldn't articulate it that cleanly.

The competitor you mentioned, do you remember the name? I'm trying to map out who else is solving this specific "write for the platform" problem vs the generic "make it sound human" crowd. Those feel like different products to me but I might be wrong.

Thanks for the push on positioning. Already tweaking the tagline.

The text detection angle is interesting but it's the wrong layer to optimize. Reader behavior is the real signal - and readers behave differently with AI content even when it's well-filtered and polished.

Google's helpful content evaluation doesn't read the text; it measures engagement patterns across your entire domain catalog. AI-homogenized content produces shorter read times, fewer shares, lower return visit rates - consistently, even when writing quality is high.

The patterns that drive engagement (unexpected perspective shifts, domain-specific detail that signals genuine experience, narrative callbacks across long articles) are hard to replicate because they emerge from how a knowledgeable person thinks through a topic, not from pattern-matching on what similar content looks like. Teaching AI to sound less like AI is harder than fixing vocabulary - it needs to reconstruct the reasoning process, not just the output.

Really liked the two-problem split someone mentioned in the comments, structural tells vs voice absence. I ran into this exact thing building the AI refinement step for DictaFlow, which cleans up dictated text after transcription. Stripping em dashes and hedging is the easy part. The hard part is making the output sound like the person who said it, not like a sanitized AI paraphrase. The approach that worked for me: let the AI fix transcription quirks like punctuation and filler words, but treat the original word choices as sacred. Don't let it rephrase anything. The model always wants to "improve" the writing, and that's exactly what makes it sound fake. Keep it on cleanup only and the voice stays intact.

X, LinkedIn, Reddit, Indie Hackers — they all have different rules.

A sentence that sounds fine on LinkedIn can feel cringe on Reddit. And most AI writing tools still make everything sound too polished, too balanced, or too “here’s the key takeaway.”

So I really like the idea of treating this as cultural adaptation, not just rewriting.

One thought: maybe don’t position it as another AI writing tool. That market feels crowded. The sharper positioning might be something like: “turn your real thoughts into platform-native English posts without sounding like AI.”

That feels more specific and more painful.

Btw, I saw a Reddit-focused writing tool here a few hours ago, so I think this “platform-native writing” pain is real. The challenge is probably how to make your angle feel different enough.

The em dash thing is so real. I ran into the exact same problem building Rowdrop.
Ended up hard-coding a regex strip on the backend because you genuinely cannot
trust the model to follow instructions. Solid build.

Spot on about the two-problem split. Structural tells are the easy part, which is why I focused there first. Voice injection is the harder problem and you framed it perfectly. Right now VoiceBridge takes a raw thought plus platform context. I am experimenting with a hot-take field where you write one opinionated sentence and the model builds around it. Early results are promising. Thanks for genuinely helpful framing.

The 'stop sounding like AI' problem is actually two separate problems:

1. Structural tells - em dashes, unnecessary hedging, bullet points for everything. Filterable.

2. Voice absence - no specific opinion, no concrete story, no sharp edge. Not filterable because filtering cannot add what was never there.

The second problem is why most AI content feels hollow even after you strip the verbal tics. The words are fine but there is no perspective behind them.

The writers I have seen use AI well do not fight this - they front-load their own specific take, then use AI to build out the structure. 'Here is my actual opinion on X: [specific, opinionated take]. Now help me write a post that builds the argument.' That forces the model to work inside someone's worldview rather than defaulting to the median internet opinion.

The filtering tool addresses the structural tells well. Curious whether you have any hooks for injecting voice/perspective, or if that is intentionally out of scope.

The strongest part here is “cultural adaptation,” not AI writing. Most writing tools promise better posts, but your wedge is much sharper: helping non-native founders turn real thoughts into platform-native English without sounding translated or AI-generated.

The three-layer system is also a good trust signal. Anti-fabrication rules, cultural phrase mapping, and AI-pattern cleanup make this feel more serious than a generic rewrite tool. I’d make that the core positioning: not “write better social posts,” but “keep your real voice while crossing language and platform barriers.”

One thing I’d watch is the VoiceBridge name. It explains the translation/voice idea, but it may still sound like a language tool. If this becomes a broader trust layer for founder communication, content, and global positioning, Beryxa .com would give it a cleaner SaaS-style brand than a feature-like name tied only to voice bridging.