Spotlight: Prose Coach — teaches your LLM to stop writing like an LLM
Most AI writing detectors don't catch bad words, they catch uniformity: same sentence length, same safe distance from anything that sounds like an actual person. Ask an LLM to write "professionally" and it reaches for the median of its training data, which is exactly that uniform, hedge-everything voice. Prose Coach's read on the problem is that you can't fix this by editing the output, because the flatness is baked in during generation. So instead of a paraphrase-after-the-fact tool, it's a text file you paste into Custom Instructions (or a custom GPT, or a Claude project) once. It sits in the system prompt and constrains the model's choices as it writes: varied rhythm, irregular sentence structure, the small imperfections that read as effort rather than autocomplete.
The smart decision here is architectural, not cosmetic. Nearly every other "humanizer" tool works downstream: you generate a draft, then run it through a rewriter that swaps words and shuffles clauses. That's a patch on a symptom, and it usually reads exactly like what it is, a patched draft. Prose Coach moves the intervention one layer up, into the instructions the model reads before it generates a single token. Load it once and it stays active for every draft after that, no re-prompting, no copy-pasting through a second tool.
The other detail worth noting is the voice-preservation valves. Any system that suppresses "AI-sounding" patterns risks flattening whatever is actually distinctive about how you write. Prose Coach is explicit that it's targeting the model's statistically common defaults, not the user's style choices, meaning a fragment that lands or a phrase you'd actually say out loud is supposed to survive the pass. Whether that holds up in practice will vary by how heavily you already lean on the AI to think for you versus just draft for you, but it's the right thing to be worried about, and most competitors don't seem to worry about it at all.
The free version is a real free tier, not a lead-gen trap: five vocabulary habits, three structural patterns, one cleanup pass, no signup, plus a scanner on the page you can paste AI-drafted text into and see the markers counted. PRO ($39) is the fuller system: 40+ categories, content-type routing for emails versus cover letters versus blog posts versus reports, and a check on sentence rhythm and voice drift before the draft comes back to you.
Who this is for: anyone doing real volume of AI-assisted writing across formats where a canned, robotic-sounding draft actually costs you something, cover letters, grant applications, client emails, blog posts you'll publish under your name. If you're already pasting drafts through Custom Instructions or running a custom GPT, this slots in with zero workflow change.
Who this isn't for: if you're typing prompts into plain ChatGPT with no custom instructions or system prompt access, there's nothing here to load, so the tool doesn't apply. It's also not a fix for content that's actually generic in substance, ideas, or argument, not just cadence, no rhythm variation will make a hollow cover letter compelling. And it's worth being honest that "make AI text pass as human" sits in ethically gray territory depending on context; using it to sound less robotic in your own business emails is different from using it to submit AI-written work somewhere that explicitly bans it.
Where I'd guess this goes next: the obvious expansion is deeper content-type routing (legal, academic, technical docs each have their own "tells") and maybe a way to fingerprint your actual past writing so the valves have something concrete to preserve, rather than a generic notion of "your voice." That second piece would be the real differentiator if they build it.
Try Prose Coach: prose.coach
See the launch: Prose Coach on welaunch.sh