Unslop Tools Are Turning AI Voice Into a Lint Error

A Show HN crossed the front page this week pitching unslopai.com, a paid service that rewrites AI-generated text so it stops sounding AI-generated — fewer "leverages," fewer "delves," no more "in today's fast-paced world." The post itself barely registered: single-digit points, and the top comment was "fighting slop with slop." But it's not an isolated launch. It's the SaaS wrapper on something that's quietly become a tool category. There are now at least three MIT-licensed "unslop" projects on GitHub, a pip install unslop on PyPI, and marketplace listings for Claude Code, Cursor, and Windsurf. De-AI-ing prose is turning into a lint stage.

That's worth taking seriously even if this particular product isn't, because the shape of the tooling tells you something about where AI-assisted writing pipelines are heading.

The taboo list became a commons

Every one of these tools sits downstream of the same upstream artifact: Wikipedia's Signs of AI writing field guide. WikiProject AI Cleanup editors have been triaging undisclosed AI content since 2023 — over ten thousand articles by their count — and the roughly 15,000-word catalog they distilled from that triage got broad coverage last fall as the best AI-writing detection guide anyone had produced. One of the open-source unslop projects, theclaymethod/unslop, literally maintains a taboo-phrases file with automated sync against it.

This matters because it means the lexical layer of the problem is now commoditized. Detecting "let's dive in," "I hope this helps," stacked hedges, negative parallelism ("it's not just X, it's Y"), and em-dash pileups is a solved, deterministic, regex-shaped problem, and the pattern list is effectively public infrastructure maintained by Wikipedia editors. Nobody should be paying $19 a month for that layer, which is the awkward position the SaaS entrants are in.

The hard part isn't the words

The interesting engineering is one level up. MohamedAbdallah-14/unslop — the more polished of the open-source pair, around a hundred stars — runs a layered pipeline: a deterministic regex pass for the known tells, then a structural pass that deliberately reintroduces sentence-length variance ("burstiness"), then an optional LLM pass for semantic rewrites, with detector scoring as a feedback loop. The theclaymethod project goes further into what it calls silhouette tells: outline-following structure, recap codas, the self-Q&A rhythm — the sense that a piece was assembled from a bullet list rather than argued.

Those structural and silhouette tells are the real fingerprint. Individual words aren't; actual humans do reach for "leverage," just less often. AI voice is a distributional property, which is why stripping the vocabulary while leaving the rhythm produces the uncanny result one HN commenter nailed as having "a whiff of automatic translation." Credit where due: theclaymethod's own published evals label the current version "no-ship" — better recall than baseline, not yet meeting its precision and safety bars. That's an unusually honest self-assessment for this space, and more trustworthy than the competing repo's self-reported 92% slop reduction on its own nine-fixture suite, which you should read as a smoke test, not a benchmark.

Where this fits in a real pipeline

The obvious objection — raised in the HN thread — is that this is a prompt problem: tell the model to write plainly and you're done. That's half right and misses why the linter form factor wins anyway. Style instructions in a system prompt fail probabilistically and invisibly; the same prompt drifts across model versions, and nobody diffs the drift. A deterministic post-pass is the opposite: it's reviewable, versionable, and enforceable. You can run it as a pre-commit hook on docs/**/*.md, fail CI when slop density crosses a threshold, and read the rewrite as a diff before it merges. The open-source tools already ship in exactly this shape — Claude Code plugin, Cursor rule files, standalone CLI — and both are careful to leave code blocks, URLs, tables, and frontmatter byte-identical, which is the property that makes them safe to point at a docs tree.

If your team ships AI-assisted changelogs, release notes, or documentation, this is worth an afternoon: it's the difference between "we asked the model nicely" and "we have a style gate." Start with the deterministic pass only. The LLM rewrite modes reintroduce the meaning-drift risk you were trying to engineer out.

The treadmill you shouldn't get on

There's a fork in this road, and the tools themselves mark it. One of the repos ships an explicit "anti-detector" mode tuned to beat AI-text classifiers. That's a different product with a different customer, and it's a treadmill: every tell these tools strip at scale stops being a tell, detectors retrain on the residue, and the cycle repeats. Wikipedia's editors flag this themselves — their guide is "descriptive, not prescriptive," a snapshot of current model behavior, not a stable spec. Any tool that hard-codes today's tells has a built-in decay rate.

Unslop-as-evasion loses that race eventually. Unslop-as-linter doesn't need to win it — its job isn't to fool a classifier, it's to stop your own docs from reading like a press release, and that bar doesn't move when the models do. The SaaS versions will struggle because their entire value is a phrase list the commons now maintains for free. The skill-and-hook versions will stick around for the same reason prettier did: once prose quality is a diffable, CI-enforceable check instead of a vibe, teams don't go back.