Your text doesn't sound like AI.
A pattern-matching filter for writers who use AI but want to keep their voice. Paste your draft. See the slop. Fix it.
No signup to try. Free forever for the basic web filter.
You can hear it now. So can your readers.
Tapestries.
Delving.
Navigating landscapes.
Em-dashes everywhere.
Sentences that "underscore the importance" of nothing.
AI-slop is a vocabulary, a rhythm, a structural fingerprint. Once you notice the patterns, you can't unnotice them. Your readers noticed last quarter.
If your writing leans on AI drafts, the slop slips in. Your voice gets diluted. Trust drops with every paragraph that reads like ChatGPT.
- ▸ You spend hours editing AI drafts, still miss the obvious tells
- ▸ Substack subscribers churn after 2-3 issues that feel generated
- ▸ Your LinkedIn posts blur into the same template everyone else copy-pasted
- ▸ Clients ask if you used AI, even when you wrote it yourself
- ▸ You stopped recognising your own voice in your own newsletter
The fix is not "stop using AI". The fix is catching the patterns before publish.
Run my draft through the filter →
Paste. Flag. Fix.
-
1.
Paste your draft into the web filter (no account needed).
-
2.
See every flagged pattern highlighted inline: banned phrases, em-dash density, parallel triples, generic openers, weak verb chains.
-
3.
Rewrite with the suggestion, or override if your style genuinely uses the pattern.
What's under the hood
- ▸ 50+ regex patterns built from a one-year archive of AI output
- ▸ No LLM. No "AI-detection magic". Pure pattern matching = transparent + fast
- ▸ Methodology open on GitHub. Add your own rules locally if you want.
- ▸ Vietnamese support: tier 1 patterns for VN content writers (beta)
Built by a linguist. Not a chatbot.
Komaru is a solo project by Komaru (founder pseudonym), a linguist based in Hà Nội, Vietnam.
Komaru runs on a one-year archive of Claude-Native vault patterns: 70+ skills built solo, all engine rules open on GitHub. The anti-slop framework comes from spotting AI patterns in linguistics work first, then porting the rules to general content.
Solo founder. No VC. No team. All metrics public on LinkedIn each week.
Three ways in. Start free.
Filter Pro lifetime + 50 banned phrases doc + 30 voice-preserving prompts + Substack growth guide.
Pays back in 4 months vs Pro monthly.
Get the Bundle →Premium phrases library. Custom rule sets per project. API access for your editor or CMS.
Cancel anytime. First month free.
Start Pro free →Web filter. 50+ patterns. No signup needed. Save your results with email.
No credit card. 5,000 chars/check, 10/day.
Run my draft through the filter →- 30-day refund. Email reply within 24h.
- Async-only support: email only, no calls or DMs.
- Built solo in Hà Nội. No VC. Public weekly metrics.
FAQ
Why not LLM-based detection?
Pattern matching is transparent. You can see why a phrase got flagged. LLM-based "AI detection" is a black box and produces false positives that erode trust. Komaru shows the rule that fired every time.
How is this different from Grammarly or Hemingway?
Grammarly checks grammar. Hemingway checks readability. Neither flags AI-slop vocabulary or structural patterns. Komaru is built specifically for the patterns that signal "this came from a chatbot".
Vietnamese support?
Tier 1 patterns for VN content writers in beta. Full coverage planned by Q3 2026. The English engine ships first.
Free tier limits?
5,000 characters per check. 10 checks per day per IP. Save results with email signup. No credit card.
Refund policy?
30 days, no questions, email reply within 24h. Bundle and Pro both covered.
Open source?
Engine rules visible on GitHub. Premium phrases library and API are closed-source under Pro. Methodology is open even when code is not.
What if my natural style uses some flagged phrases?
Override per phrase, or add a personal allowlist in Pro. The filter flags patterns; you decide what's slop and what's voice.
Will Tool A stay rule-based?
Yes. Adding LLM features defeats the "transparent + fast + cheap to run" promise. Premium features expand the rule library and add per-project custom rules. Engine stays regex.