- Published on
Linting my blog for AI-sounding prose: what worked, what I skipped

- Authors

- Name
- Nadim Tuhin
- @nadimtuhin
On this page
I write a lot of my posts with an AI model in the loop. So I had a nagging question: how much generic phrasing had leaked in?
My goal was a list of sentences to rewrite, with no detector score involved. Here's what I ran, what it found in my own posts, and the check that now runs on every commit.
The tools I looked at and skipped
Search for "make AI text sound human" and the first page is a pile of paid tools. I read through the list and set most of them aside.

AI detectors (GPTZero, Originality, Turnitin, Quetext, Smodin). Independent tests put them well below their own marketing numbers, and they flag real human writing. A Stanford study found they labelled most essays by non-native English writers as AI. I'm a Bangladeshi engineer writing in English, so a detector could easily flag my own sentences. And nobody who hires me will run one.
Paid humanizers (HixBypass, StealthWriter, WriteHuman, Phrasly and similar). They rewrite text to dodge detectors. Mostly they shuffle words, and the output reads worse. The "tested and ranked" pages that praise them are often affiliate sites, so I gave those little weight.
Cheap shared-login seats. Some lists sell premium tools for a few dollars a month. They share one account between many people, which breaks the vendor's terms. You'd also be signing in from the same browser you use for your real accounts.
What I used instead
slopless is an npm package built on textlint. It runs locally, makes no model calls, and prints JSON. It checks for the contrast-reframe pattern, stock phrases such as "tip of the iceberg", and filler words. Its dependencies are textlint and a few text libraries, and it has no postinstall script.
Harper is an offline grammar checker with editor plugins. It suits a second-language writer like me because it runs on your machine.
Vale with an AI-tells style is the heavier option if you want rules you can edit. I didn't need it yet.
What it found in 31 posts
One snag first: slopless only reads .md files, and my posts are .mdx. So I copied each post to a temp folder, removed the frontmatter and code blocks, and ran it there. Across 31 posts (evaluating ~42,000 words in under 1.2s), it reported 37 findings.
Some were noise. It flagged curly quotes in my poems and title case on product names. I ignored those.
The rest were real. This is the worst offender, from a post about software in 2026:

- "What changed wasn't the AI. It was my mental model." became "What changed was my mental model."
- "Adapting to coding agents is just the tip of the iceberg." became "Adapting to coding agents is the easy 10%."
- "Here's what I'm actually using in 2026:" lost the word "actually".
Across 13 posts I made 25 edits. The word "actually" went from five uses to one in that post. The leftovers were false positives and one sentence about webhook delivery guarantees, which I reworded anyway so the check would pass cleanly.
I didn't try to reach zero. Fixing every "not X, but Y" sentence would flatten how I write, and some contrasts are the point.
The check that runs on commit
A one-off clean-up fades. A check on every commit sticks.


scripts/lint-prose.mjs does the same copy-and-strip step, runs slopless, drops the noisy rules, skips the poems, and prints the file, line number and sentence. lint-staged runs it on any staged post under data/blog/. If it finds something, the commit stops.
$ node scripts/lint-prose.mjs data/blog/test-post.mdx
test-post.mdx:14 contrast-reframe
What changed wasn't the AI. It was my mental model.
1 prose finding(s). Rewrite the sentence, or skip with: git commit --no-verify
To test it, I staged a post with a deliberate contrast-reframe sentence added. The commit was blocked in ~450ms, and the output named the line. I removed the test line afterwards and confirmed the file was clean.
git commit --no-verify skips the check when I disagree with it. A hook that can't be bypassed gets deleted.
Lessons I'd keep
- Start with a list of sentences to rewrite, not a score.
- Ignore the rules that flag your own style on purpose.
- Fix the clear cases and stop when only false positives are left.
- Put the check in a hook so new posts get it too.
The whole thing is one script and one dev dependency. The paid tools I skipped run from a few dollars to about fifteen a month, and I don't think they'd have told me anything this didn't.
Related posts
- Hermes Agent: nightly cron jobs and skills learned from corrections
Hermes runs nightly CLI updates, a security PR scan every 6 hours and a 30-minute autoheal job. Every correction I make becomes a skill file. Hermes drafts, I send.
- Go berserk: one night with Grok Bot, Orca and every gate still on
My Grok Bot usage reset, so I told my AI teammates to go berserk overnight. They set up Orca, fixed a Windows PC and refreshed my blog, and every order kept the review gates on.
- 9agent: run Claude Code, Pi or Hermes on any model with one command
A small launcher that points Claude Code, Pi or Hermes at a 9Router gateway, lets me pick any model, and can run the agent in Docker. It never rewrites configs.