anti-slop-kit
A grep gate, a pattern check and a critic loop that make AI-assisted writing read like its author.
Safety scan
A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.
Badge for your README
Markdown
[](/listings/mavrin-xv--anti-slop-kit)HTML
<a href="/listings/mavrin-xv--anti-slop-kit"><img src="/api/listings/mavrin-xv--anti-slop-kit/badge.svg" alt="findsafeskills safety rating" /></a>Install
Install instructions from the README
As written by the repo's author. findsafeskills has not verified these steps.
Everything runs offline on Fernhill Bikes, an invented repair shop.
python3 skills/no-ai-slop/scripts/slop_grep.py examples/fernhill-bikes/draft-sloppy.md
== examples/fernhill-bikes/draft-sloppy.md [long copy] FAIL banned:3 game-changer In today's world, riders want a seamless experience — and tubeless FAIL banned:3 seamless In today's world, riders want a seamless experience — and tubeless tyr FAIL phrases:3 In today's world In today's world, riders want a seamless experience — and tub FAIL banned:5 robust It's not just about fewer flats. It's about confidence. When it comes to c FAIL phrases:5 When it comes to It's not just about fewer flats. It's about confidence. When it FAIL cluster:7 actually, honestly, just Honestly, the conversion is actually simpler than most FAIL em-dash:0 4 on latin lines, budget 2 (long copy) warn adverbs:5 just ... warn soft:13 Ultimately Ultimately, the best tyre is the one that gets you home. score 0/100 (214.8 weighted hits per 1,000 words, 128 words)
Exit 1. The same post rewritten in the shop's voice:
== examples/fernhill-bikes/draft-clean.md [long copy] warn no-source:3 241 Ticket 241: a courier with four flats in a month, all thorns, all on the s no hard hits: review the warnings, then read for the patterns score 84/100 (5.8 weighted hits per 1,000 words, 173 words)
Exit 0. The warning is a number with no source next to it; here it is the shop's own ticket, so it stays.
python3 skills/voice-extractor/scripts/voice_stats.py examples/fernhill-bikes/samples/*.md --draft examples/fernhill-bikes/draft-sloppy.md
| dashes_per_1k | 0.0-0.0 | 0.0 | 30.3 | DRIFT | | question_share | 0.0-0.0 | 0.0 | 0.14 | DRIFT | | first_person_per_100 | 0.8-3.3 | 2.02 | 0.0 | DRIFT | ... 7 drift flags
The clean draft gets 0 drift flags. The detect report, the critic loop run and the voice check for both drafts are in [examples/fernhill-bikes/](examples/fernhill-bikes/).
What was scanned
The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.
.claude-plugin/marketplace.json · Manifest · 15 lines
- {
- "name": "anti-slop-kit",
- "plugins": [
- {
- "name": "anti-slop-kit",
- "source": "./",
- "description": "Agent skills that make AI-assisted writing read like its author: a grep gate that exits 1 on banned words, empty phrases, dash overuse and clusters of soft tells, a by-hand pattern check, a five-dimension critic loop with a retry cap, a voice profile from real samples and a drift check against it. Optional Ukrainian and Russian packs.",
- "version": "0.1.0"
- }
- ],
- "owner": {
- "name": "Artem Mavrin",
- "url": "https://github.com/mavrin-xv"
- }
- }
README.md · README · 162 lines
- 
- # anti-slop-kit
- **Four agent skills that make AI-assisted writing read like the person who signs it: a grep gate that exits 1 on banned words and clusters, a by-hand pattern check for what no script sees, a five-dimension critic loop with a retry cap, and a voice profile your drafts are checked against.**
- Drafts written with a model share the same tells: the same twenty words, the same "it's not X, it's Y", a dash in every sentence, a moral at the end. Readers spot them in one line and stop trusting the rest.
- [](LICENSE)
- 
- 
- 
- 
- ```bash
- git clone https://github.com/mavrin-xv/anti-slop-kit && cd anti-slop-kit
- ./install.sh # or: ./install.sh --project
- ```
- <!-- install-notes:start -->
- Other ways in: as a Claude Code plugin, run `/plugin marketplace add mavrin-xv/anti-slop-kit` and then `/plugin install anti-slop-kit@anti-slop-kit`. For Codex or another agent that reads `SKILL.md`, copy the folders under `skills/` into its skills directory. On Windows, run `install.sh` from Git Bash or WSL and type `py -3` where these docs say `python3`.
- <!-- install-notes:end -->
- Then tell your agent: `check this draft for slop` or `does this sound like me?` or `score this post before I publish it`
- ---
- ## What it solves
- | Pain | What the pack does |
- |---|---|
- | The same AI words in every draft | `slop_grep.py` checks a word list, empty phrases and context-aware patterns (a "10x position" in finance is not flagged; a marketing use of the same verb is) and exits 1 on a hard hit |
- | Dashes everywhere | A dash budget per format: none in short copy, one per 150 words in long copy, not counted in scripts read aloud |
- | One soft word is fine, five in a paragraph is a tell | Adverbs are warnings alone; three different ones in one paragraph are a hard `cluster` hit |
- | The text passes the word list and still reads as machine-written | The reading half: 18 named patterns plus four structural tells, and a rule that any three together fail |
- | "Is this better or worse than last time?" | A score out of 100 from weighted hits per 1,000 words, where repeats cost more than variety, and `--min-score` as a gate |
- | A draft that feels done but is not | `critic-loop`: five dimensions with fixed bars, the lowest fixed line by line, two retries, then a human decides |
- | "That does not sound like me" | `voice-extractor` builds a seven-field profile from 5 to 30 samples with quoted evidence; `voice-check` scores six dimensions of drift and quotes each one |
- | Rhythm that drifted, but nobody can say how | `voice_stats.py` measures sentence length, spread, fragments, dashes, questions and first person, and flags draft metrics outside the samples' range |
- | Ukrainian or Russian copy | Optional packs, loaded automatically for Cyrillic text; the long dash there is grammar and is never counted |
- ## What's inside
- | Skill | What it does | What to say |
- |---|---|---|
- | `no-ai-slop` | Detect (name and quote) or edit (minimum rewrite plus What changed); the grep gate, the pattern half, an eval checklist, a copy layer for selling surfaces, UA and RU packs | "check this for slop", "make this sound less like AI", "audit this before it ships" |
- | `critic-loop` | Hook, facts, economy, reader path, slop-free; surgical fixes, at most two retries | "score this draft", "critique this post", "tighten this before I publish" |
- | `voice-extractor` | Seven fields with evidence and confidence, measured rhythm, a falsifiability test, a three-round test loop with the owner | "capture my writing style", "build a voice profile from these posts" |
- | `voice-check` | Six drift dimensions against the profile, measured drift, a slop floor, drift vs evolution, a five-point fingerprint | "does this sound like me", "check the ghostwriter's draft" |
- ## Quick start
- Everything runs offline on Fernhill Bikes, an invented repair shop.
- ```bash
- python3 skills/no-ai-slop/scripts/slop_grep.py examples/fernhill-bikes/draft-sloppy.md
- ```
- ```
- == examples/fernhill-bikes/draft-sloppy.md [long copy]
- FAIL banned:3 game-changer In today's world, riders want a seamless experience — and tubeless
- FAIL banned:3 seamless In today's world, riders want a seamless experience — and tubeless tyr
- FAIL phrases:3 In today's world In today's world, riders want a seamless experience — and tub
- FAIL banned:5 robust It's not just about fewer flats. It's about confidence. When it comes to c
- FAIL phrases:5 When it comes to It's not just about fewer flats. It's about confidence. When it
- FAIL cluster:7 actually, honestly, just Honestly, the conversion is actually simpler than most
- FAIL em-dash:0 4 on latin lines, budget 2 (long copy)
- warn adverbs:5 just ...
- warn soft:13 Ultimately Ultimately, the best tyre is the one that gets you home.
- score 0/100 (214.8 weighted hits per 1,000 words, 128 words)
- ```
- Exit 1. The same post rewritten in the shop's voice:
- ```
- == examples/fernhill-bikes/draft-clean.md [long copy]
- warn no-source:3 241 Ticket 241: a courier with four flats in a month, all thorns, all on the s
- no hard hits: review the warnings, then read for the patterns
- score 84/100 (5.8 weighted hits per 1,000 words, 173 words)
- ```
- Exit 0. The warning is a number with no source next to it; here it is the shop's own ticket, so it stays.
- ```bash
- python3 skills/voice-extractor/scripts/voice_stats.py examples/fernhill-bikes/samples/*.md --draft examples/fernhill-bikes/draft-sloppy.md
- ```
- ```
- | dashes_per_1k | 0.0-0.0 | 0.0 | 30.3 | DRIFT |
- | question_share | 0.0-0.0 | 0.0 | 0.14 | DRIFT |
- | first_person_per_100 | 0.8-3.3 | 2.02 | 0.0 | DRIFT |
- ...
- 7 drift flags
- ```
- The clean draft gets 0 drift flags. The detect report, the critic loop run and the voice check for both drafts are in [examples/fernhill-bikes/](examples/fernhill-bikes/).
- ## How it works
- ```
- samples (5-30) --> voice-extractor --> voice-profile.md
- |
- draft --> no-ai-slop -----------------> critic-loop ------------> voice-check --> ship
- 1. slop_grep.py exit 0? 5 dimensions, bars 6 drift dims
- 2. pattern half < 3 tells? fix lowest, retry <= 2 voice_stats --draft
- 3. edit + eval.md else hand back slop floor exit 0
- ```
- The script is a floor, not a verdict. It catches the vocabulary and the counting; the reading half catches shape, which is where most of the tells live. Every later skill re-runs the gate, so no score can carry an exit-1 draft through.
- ## Configure
- | Setting | Where | Default |
- |---|---|---|
- | Word lists | `skills/no-ai-slop/references/slop-tokens.txt`, `references/packs/*.txt`; one entry per line, `re: ` for a regex | English core; `ua`, `ru` packs |
- | Packs | `--pack auto`, `none`, or a list such as `ua` | auto: all packs when 30% of words are Cyrillic |
- | Your own list | `--tokens my-list.txt` | the core file |
- | Mode | `--mode short`, `long`, `spoken`, `auto` | auto from the file name: `*.spoken.txt` spoken, `slides.md` short, else long |
- | Dash budget | `EM_BUDGET` at the top of `slop_grep.py` | short 0, long max(2, words/150), spoken not counted |
- | Cluster size | `CLUSTER_AT` in `slop_grep.py` | 3 different soft words in one paragraph |
- | Score gate | `--min-score N` | off |
- | Critic bars | `skills/critic-loop/SKILL.md` | hook 8 (first line 9), facts 8, economy 8, reader path 3/5 |
- | Drift range | `voice_stats.py` | sample min to max, widened by a quarter of the range |
- Install without the script: copy the folders in `skills/` to `~/.claude/skills/` (or your agent's skills folder).
- ## Honest limits
- - The word list catches vocabulary, not authorship. A clean exit means no listed tell, not "a human wrote this". The skills never claim to detect who wrote a text.
- - The score is a trend number on one draft over its edits. Do not compare scores across formats or authors.
- - The reading half depends on the model reading carefully. It is a checklist, not a detector.
- - Word lists age as models change their habits. Expect to add entries; `tests/fixtures/` is where each one gets a test.
- - `voice_stats.py` splits sentences with a regex. Abbreviations and odd punctuation shift the counts a little; read drift rows next to the samples.
- - The Ukrainian and Russian packs are smaller than the English core and were checked on a few hand-written texts only.
- - All example businesses and texts are invented. Tested on Windows with Git Bash and Python 3.12; not tested on macOS or Linux, and `install.sh` was not run end to end.
- ## Credits
- - Based on / inspired by https://github.com/petergyang/no-ai-slop (MIT): the two jobs, the editing principles, the patterns to cut and the eval checklist, lightly trimmed
- - Based on / inspired by https://github.com/boringmarketer/boring-marketing-skills (MIT): the structural tells and the density rule, the seven voice fields and the test loop, the six-dimension drift check, extra word-list entries
- - Based on / inspired by https://github.com/eric-tramel/slop-guard (MIT): the density score with exponential decay and the repeat penalty
- - Based on / inspired by https://github.com/conorbronsdon/avoid-ai-writing (MIT): severity tiers where soft tells count only in a cluster, and evals that test the checker
- - Based on / inspired by https://github.com/blader/humanizer (MIT): strong patterns that fail on one sighting, and the draft, audit, final pass
- - Based on / inspired by https://github.com/hardikpandya/stop-slop (MIT): scoring with a pass bar before a draft ships
- - Based on / inspired by https://github.com/jdkato/voices (MIT): fixtures that must fire and fixtures that must stay silent
- - Based on / inspired by https://github.com/sam-paech/slop-forensics (MIT) and https://github.com/sam-paech/antislop-sampler (Apache-2.0): profiling over-used words against a baseline, the idea behind the fingerprint candidates in `voice_stats.py`
- - Based on / inspired by https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing (Wikipedia, CC BY-SA): the catalogue of tells by section; ideas restated, no text copied
- ## Contributing
- See [CONTRIBUTING.md](CONTRIBUTING.md). Notes for coding agents are in [AGENTS.md](AGENTS.md).
- <!-- author:start -->
- ## Author
- Built and maintained by [Artem Mavrin](https://github.com/mavrin-xv). This repo is one of a set of agent skill packs for marketing and operations work. Each one installs on its own; the full list is in [agent-skill-packs](https://github.com/mavrin-xv/agent-skill-packs). If a pack saved you an afternoon, a star helps other people find it.
- <!-- author:end -->
- ## License
- MIT. See [LICENSE](LICENSE).
Ratings and comments
Sign in to rate this skill.
Sign inFree to use. If it’s helped, pass it on.