The only linter built for what LLMs produce — not for what humans get wrong.
A toolkit for cleaning the LLM out of LLM-assisted writing. Catches the tells in English and Russian, in under fifty milliseconds, with no model call. Twelve wrappers compose on top for the genres LLMs handle worst: landing copy, release notes, fiction, cold email, image and video prompts.
No model call, ~50 ms on a 5K-word file. Each hit carries category and severity. Same diagnosis on every machine — deterministic by design.
Every LLM-assisted draft I shipped was starting to read like every other LLM-assisted draft.
The cost of writing with LLMs isn’t time — it’s that the drafts stop sounding like you. They start sounding like Claude’s version of you. Once you notice, you can’t unsee it: “it’s important to note that…”, “delve into the rich tapestry of…”, “in today’s fast-paced world…” — every LinkedIn post, every product update, every investor email wearing the same set of tells, all of them blurring into one voice. Readers spot it before you do.
So I built a tool that strips those tells. Then I extended it. Now it is one offline linter and seventeen agent skills, MIT-licensed, with one curl to install. The linter is pure Python — fifty milliseconds, deterministic, no model call, runs in any CI or on the command line. The skills are plain markdown specs; Claude Code loads them natively today, and any agent that reads markdown skills can use them tomorrow.
Composed in layers. Each wrapper has a sharp `description:` field so Claude Code picks the right one without colliding with the others.
The base `writer` linter plus five wrappers that shape the prose for a specific genre or register.
writerbase regex linter, 28 categories (EN + RU)viral-texthooks, numbered points, NLP question, CTAprose-editfiction (Pelevin / Manson voice vectors)essay-writenon-fic longread, Manson coda, V/H/P markerstone-shifterregister shifts (casual ↔ academic)pelevin-digressionopinionated voice insertsNewer wrappers (v1.6+). Each ships banned-pattern catalogues plus before/after calibration pairs.
landing-copyhero, features, pricing, SEO, adsrelease-notesKeep-a-Changelog, per-audience tonerfc-writerRFCs / ADRs / Tech Specs (RFC 2119)microcopyerror states, empty states, buttonscold-email5-block, under 120 words, banned ceremonyimage-promptMJ / DALL-E / Flux (6-part formula)video-promptKling / Veo / Sora (beat structure)Read-only — report and exit. Designed for CI gating. The meta-skill manages the collection itself.
style-checkpre-commit prose gatetranslation-syncRU↔EN↔PT-BR paritycanon-checkstory-bible consistencyskills-updatechecks GitHub and applies new releaseSkills land in `~/.claude/skills/` and are auto-discovered when Claude Code restarts. No `settings.json` edits required.
curl -fsSL https://raw.githubusercontent.com/Mikefluff/skills/main/install.sh | bashNo deps beyond bash + python3. macOS + Linux.
Updates, uninstall, and troubleshooting at docs/INSTALL.md.
New skill ideas, banned-pattern updates, and the occasional rewrite teardown. One email per release. Unsubscribe with a click.
A paragraph of landing-page slop on the left. The same paragraph after running `landing-copy` on the right. The linter report and the named changes below.
We're excited to announce our revolutionary, game-changing platform — built by an industry-leading team using state-of-the-art technology.
Save time, boost productivity, and get more done with our world-class solution. Click here to learn more.Cut PR review-time from two days to four hours on the teams that adopted the workflow last quarter. Concrete numbers per team in the case studies.
See pricing — free for 30 days.writer-lint: neuroslop suspected 14 hits By severity: blocker 0 caution 14 nit 0 Top categories: MARKETING_HYPE 6 VAGUE_BENEFIT 3 EMPTY_CTA 2 WEAK_OPENER 1 NE_X_A_Y 1 CORPORATE 1
Same word budget. The before sounds like every other LLM-generated landing. The after sounds like a product.
Twelve scenarios grouped by domain. Each row points to the skill — or chain of skills — that handles it.
prose-edit + canon-checkessay-writeviral-texttone-shifterlanding-copy + microcopyrelease-notescold-emailrfc-writerimage-promptvideo-promptstyle-checktranslation-syncFull taxonomy and fourteen named recipes at docs/SKILL-INDEX.md · docs/COMPOSING.md
Drafts stop coming back with "this sounds AI-generated" comments. First-pass cleanup goes from an hour to ten minutes. The voice you started writing in stays your voice instead of converging on Claude’s default. If the page you ship is your judgment instead of your model’s, the rest of the team starts treating it that way too.
Grammarly catches grammar. LanguageTool catches typos and basic style. Hemingway catches sentence length and adverbs. None of them catch the specific patterns LLMs produce — "navigate the complexities", "rich tapestry", balance hedges, AI-style triplets, marketing hype ("revolutionary", "world-class"), excitement preambles ("we’re excited to announce"). This is purpose-built for the LLM-output problem, which is a different shape than human-error catching.
For pre-commit and CI gating, regex wins: ~50 ms vs 5-10 seconds, no API cost, deterministic, no LLM dependence. For nuanced rewriting, the wrappers do use the LLM — that is where context matters. The linter is the floor; the wrappers are the ceiling.
English coverage is now first-class — all 28 linter categories have EN patterns (the marketing-specific set added in v1.8 is EN-led). On a synthetic English neuroslop fixture, the linter fires 54 hits across 18 categories. Fourteen of seventeen skills support both languages. The three Russian-only skills are author- and language-specific.
Yes. `python3 writer/scripts/lint.py file.md` works standalone. The wrappers require Claude Code (they are LLM-driven by design). A Docker image exposes the linter for CI without any Python install on the host.
MIT-licensed, no external deps, tested on real prose for over a year. Best thirty seconds you’ll spend this week.
The linter came out of the same habit as everything else here: name the thing precisely, then fix the named thing. It works on businesses too.
v1.9.1 · MIT · ghcr.io/mikefluff/skills · @mikefluff/skills