A lint pass for AI-sounding prose, with receipts
unslop detects and repairs formulaic AI writing while protecting facts, legal hedges, and your actual voice — and it honestly reports it doesn't pass its own bar yet.
What it does
unslop is an agent skill (plus standalone Python scanners) that finds AI writing tells — “It’s not X, it’s Y”, “Let’s dive in”, em-dash pileups, moralizing codas — and rewrites them while preserving meaning. It ships four commands: teach builds a voice profile from your own writing, cleanup flags without editing, rewrite repairs drafts, and mimic drafts in your taught voice. The scanners run on plain Python 3.8+ with no dependencies, so CI or any other agent can call them directly.
The interesting bit
The project’s most unusual feature is its honesty about its own limits: the latest independent benchmark result is published as a no-ship — it beat the baseline model on recall and repair, but missed the precision and collateral-damage bar. It also separates “engineering works” from “writing improves” into different scoreboards so passing tests can’t masquerade as product quality.
Key highlights
- Three detection layers: literal phrase triggers with false-positive guards, 36 structural patterns (sentence burstiness, signpost density, listicle counts), and a silhouette layer scoring how ideas are arranged —
callback_contentbeing the strongest single tell. - Contextual gating over blunt word bans: “may cause drowsiness” and “notwithstanding anything to the contrary” are content, not filler, and survive rewrites.
- Magnitude-aware fact preservation —
$47.3Mcan’t quietly become$47.3 billion. - Quoted spans and code fences are masked, so documentation about bad writing doesn’t flag itself.
- Genre carve-outs: bold-label lists are fine in reference docs, staccato lines are fine in social copy.
Caveats
- English only; non-English input gets a clean refusal rather than a mangled edit.
- The core benchmark is currently a no-ship — precision, damage, and whole-document safety fell short of the bar, with failure modes documented in
evals/CORE-RESULTS.md. - A clean scan proves nothing about authorship; the README is explicit that neither scanners nor familiar phrases can attribute who wrote a text.
Verdict
Worth a look if you publish AI-assisted drafts and want a rigorous, self-critical de-slopper rather than a find-and-replace of “delve”. Skip it if you need non-English support or expect a tool that’s already cleared its own quality bar — it isn’t, and it says so.
Frequently asked
- What is theclaymethod/unslop?
- unslop detects and repairs formulaic AI writing while protecting facts, legal hedges, and your actual voice — and it honestly reports it doesn't pass its own bar yet.
- Is unslop open source?
- Yes — theclaymethod/unslop is an open-source project tracked on heatdrop.
- What language is unslop written in?
- theclaymethod/unslop is primarily written in Python.
- How popular is unslop?
- theclaymethod/unslop has 1.2k stars on GitHub.
- Where can I find unslop?
- theclaymethod/unslop is on GitHub at https://github.com/theclaymethod/unslop.