Skip to main content

dislop — Catch AI slop before you send it

简体中文 · Try it at dislop.com · npm · PyPI · Awesome Anti-Slop · Roadmap · X (@dislop_ai) · Apache-2.0

dislop is a dedicated Linter for AI Slop. It combines vetted community anti-slop rules, corpus findings, and 87 curated tools so you (and your agent) can catch and fix AI-sounding text and template layouts before sharing docs, web pages, or slides. Connects in one command to coding agents (Codex, Claude Code, Antigravity, Cursor), office agents (Claude Cowork, WorkBuddy, Qwen), and editors, and can run completely offline locally.

Here is a short launch note straight from a model:

Certainly! Here is the launch note. It's worth noting that exports are a game-changer, finishing in up to 12 minutes for workspaces above 500,000 rows.

dislop check launch-note.md flags three things:

[ERROR]   launch-note.md:1:1   chatbot-artifact           "Certainly! Here is"
[WARNING] launch-note.md:1:37  en-throat-clearing-opener  "It's worth noting that"
[INFO]    launch-note.md:1:75  en-cliche-lexicon          "game-changer"

If you rewrite it into a second-order AI slogan (This isn't just an upgrade — it's a new standard.), running dislop check catches that too. Clean it up into plain prose and dislop check exits 0.

dislop checking a launch note across 3 states: flagging chatbot artifact and filler, catching a second-order slogan, and passing with exit 0

That's it. It inspects Markdown, plain text, HTML/CSS/JSX, Word (.docx), and PowerPoint (.pptx). The CLI and local plugins run completely offline with zero uploads. They share the same 7 built-in packs (42 rules), 5-axis rhythm telemetry, offline style extraction, extended pattern probes (--patterns), bundled 14.6 KB Pocket ML (--ml), and hardware-aware optional local models (dislop model). At dislop.com, the browser workbench switches between local rule checks and Cloud MCP mode, and you can connect Hosted Cloud MCP (https://dislop.com/mcp) whenever no local shell is available.

Quick start

Node.js 20+ (published on npm, zero dependencies, all 7 packs bundled offline)

# Check the current folder (exits 1 on core AI artifact errors; prints cliché warnings)
npx --yes dislop check .

# Auto-route domain pack and enable 5-axis telemetry, 14.6 KB Pocket ML, and extended pattern probes
npx --yes dislop check draft.md --pack auto --profile --ml --patterns

# Extract a deterministic personal voice profile offline or inspect hardware for optional local models
npx --yes dislop voice my-notes.md
npx --yes dislop model

# Check only lines changed against main and emit GitHub PR annotations
npx --yes dislop check --changed origin/main --format github

# Explain a rule's check question and before / overcorrected / after examples
npx --yes dislop explain negation-pivot-frame

Python 3.10+ (published on PyPI, zero dependencies)

# Run directly via uvx (zero install) or install via pip
uvx dislop check quarterly-review.pptx
pip install dislop && dislop check .

dislop CLI offline check failing with exit code 1 on draft and passing with exit code 0 after cleanup

AI artifacts vs. style warnings

Auto-rewriters often trade one AI cliché for another or flatten your voice into a generic template. Flat banned-word lists also cause false alarms on ordinary nouns (Delve debugger, wool tapestry).

dislop splits its 42 rules across 7 bundled packs into two layers:

  1. AI delivery artifacts (packs/core/, severity: error, always on): Clear leftovers from AI generation—stray chatbot greetings (Certainly! Here is..., [oaicite:1]), unfilled [insert ...] / {{var}} / [TBD] placeholders, raw Markdown asterisks pasted unrendered into .docx/.pptx, and dead <a href="#"> placeholder buttons.
  2. Built-in scenario & vertical packs (warning / info, advisory by default, 0 network downloads): business-brief, engineering-spec (PR diff parrots, shallow root causes, unowned action items), regulated-comms (unsubstantiated superlatives, unanchored "10x faster" multipliers), literary-essay, deck-and-ui, and executive-deck (process-only slide titles, 4+ consecutive identical slide skeletons). Each rule includes a check question, before / overcorrected / after examples, and automated hit/clean tests.

What it reads

Both the Node.js CLI (bin/dislop.mjs) and the Python CLI (src/dislop/) run with zero external runtime dependencies and pass the same conformance suite (tests/test_conformance.py).

What you bring Extensions What dislop catches Known boundary
Drafts & docs .md, .mdx, .txt AI artifacts + throat-clearing openers, "Not X, but Y" pivots, staccato slogan runs, bureaucratic padding, four-character triads, moralizing endings, unsourced claims, bold-lead bullet walls, em-dash overuse, half-width CJK punctuation Code blocks, inline code, URLs, blockquotes, and dislop-quote blocks are masked for style rules
Web pages & components .html, .css, .jsx, .tsx Dead <a href="#"> placeholder links, unexplained <button disabled>, purple-pink AI gradients, pill-badge walls, 3-level nested boxes, AI-Powered prompt-echo badges Static per-file inspection; runtime DOM state still needs browser tests
Word .docx Joined <w:t> runs, raw Markdown asterisks pasted from chat, and prose slop rules (para N) Reports paragraph number rather than rendered page number
PowerPoint .pptx Presentation slide order (p:sldIdLst), joined <a:t> runs, raw Markdown, slide word-walls (> 120 words / > 280 Han chars), bullet caps (> 6), <a:normAutofit> shrink (< 80%), generic titles (Overview, Background), layout monotony, and prose rules Hidden slides (show="0") are reported as [SKIP] for density checks
dislop.com offline browser workbench checking original draft, 2nd-order slop, and clean copy
Offline browser workbench in action
Auditing AI gradient, pill wall, and dead href='#' control before switching to clean 1px galley card
UI template slop audit before and after

Awesome Anti-Slop

Looking for other tools, style guides, design rules, PR policies, and papers in this space? Browse awesome/README.md (Chinese edition) or filter all 87 checked entries offline at dislop.com/awesome/ (machine-readable index.json and index.md).

Use it across agents, editors, and CI

Surface One-liner or 1-click install What you get
Any AI chat Paste: Read https://dislop.com/llms.txt and apply its anti-slop rules and checks before delivering my next draft. Zero install; reads the ~2.4 KB index, task guides (/tasks/*.md), or atomic rule cards (/rules/<id>.md)
Agent Skill (75+ hosts) npx skills add dislop-ai/dislop Installs skills/dislop/SKILL.md + 4 local reference cards in Claude Code, Codex, Cursor, Windsurf, Antigravity, and OpenCode (0 network calls)
Claude Code Plugin /plugin marketplace add dislop-ai/dislop then /plugin install dislop@dislop-ai Bundles Skill + local stdio MCP (~130 always-on tokens)
OpenAI Codex Plugin codex plugin marketplace add dislop-ai/dislop then codex plugin add dislop@dislop-ai Ships native UI card metadata (openai.yaml) and $dislop trigger
Gemini CLI Extension gemini extensions install https://github.com/dislop-ai/dislop Mounts local bin/dislop-mcp.mjs directly via gemini-extension.json
Official MCP Registry Search io.github.dislop-ai/dislop on registry.modelcontextprotocol.io Lists local npx / uvx stdio packages and remote https://dislop.com/mcp
VS Code / Cursor 1-Click MCP Install in VS Code (Local Offline) · Install in VS Code (Cloud) · Install in Cursor (Local) Opens the editor's native confirmation dialog without manual JSON edits
GitHub Action / pre-commit Reference uses: dislop-ai/dislop@main (action.yml and .pre-commit-hooks.yaml) Gates modified lines in PRs with inline annotations or SARIF 2.1.0

Manual MCP config (Claude Code, Codex, Cursor, Windsurf)

{
  "mcpServers": {
    "dislop": {
      "command": "npx",
      "args": ["--yes", "dislop", "--mcp"]
    }
  }
}

Exposed local MCP tools (7 offline tools, verbosity: "compact" by default, with negative fixtures stripped): dislop_check (supports pack including "auto", profile, ml, and patterns), dislop_voice_extract_offline, dislop_model_score, dislop_model_status, dislop_explain, dislop_packs, and dislop_catalog.

Configure it, or overrule it

Run dislop init to create .dislop.json (put private local bans in Git-ignored .dislop.local.json):

{
  "failOn": "error",
  "ignore": ["vendor/**", "dist/**"],
  "overrides": [
    { "files": ["docs/**/*.md"], "pack": "business-brief" },
    { "files": ["essays/**/*.md"], "pack": "literary-essay" },
    { "files": ["site/**/*.{html,css,svg}"], "pack": "deck-and-ui" }
  ],
  "allow": ["Delve", "Bayeux Tapestry"],
  "rules": {
    "bold-lead-list-wall": "info"
  }
}

For a narrow one-off exception, add a reason above the line (--report-unused flags stale comments):

<!-- dislop-ignore-next-line en-cliche-lexicon -- quoting vendor press release verbatim -->

Install once locally (true air-gapped / fast hooks)

# Node.js 20+ (published on npm, < 250 KB tarball, zero network after install)
npm install -g dislop
dislop check .

# Or Python 3.10+ (published on PyPI, zero external dependencies)
pip install dislop
dislop check .

Open-Source Local (Skill + CLI) vs. Hosted Cloud (dislop.com/mcp)

  • Open Source (dislop-ai/dislop, Apache-2.0): SKILL.md + bundled local ./references/*.md, Node/Python CLI (dislop check [--pack <id|auto>] [--profile] [--ml] [--patterns] [--changed [ref]] [--staged] [--format text|json|sarif|github|md], dislop patterns, dislop voice, dislop model [status|score|install|import|remove|deslop], dislop setup, dislop deslop --local, dislop explain, dislop packs, dislop catalog, dislop resources), 7-tool local stdio MCP (--mcp), all 7 built-in packs (42 rules) + 53-pattern catalog (patterns/), bundled 14.6 KB gzip (49.0 KB raw) Pocket ML ensemble (models/pocket-slop-v1.json, < 0.5 ms, 100.0% F1 / 0.0% Clean FPR on stealth-vapor-v1), Awesome Anti-Slop, and dislop-bench v2. Zero third-party packages, zero API keys, and 0 network calls during check / patterns / voice / model score so private drafts never leave your machine. For heavier local neural rewrites, dislop setup and dislop model install inspect RAM/disk budgets and let you optionally install 0.6B (~520 MB), 1.7B (~1.3 GB), or 4B (~2.5 GB) local models (or import offline .gguf/.onnx weights with SHA-256 verification) protected by an internal numeric and qualifier retention guard during automated local rewrites.
  • Hosted Cloud Remote MCP & Web AI (https://dislop.com/mcp, https://dislop.com/?lang=en#pricing): Optional closed-source cloud backend powered by a paid-tier zero-training Google Gemini endpoint (no draft logging; see privacy.md). Exposes dislop_review (1 credit, deeper cadence and prose review), dislop_deslop (2 credits, re-checked rewrite verifying numbers and qualifiers), and dislop_voice_extract (2 credits, dislop.voice/1). Includes 5 free anonymous web demo credits/day, 30 free monthly credits after email key signup, and early-access waitlists for Pro ($12/mo, 1,500 credits), PAYGO ($10 / 800 credits), and Team ($29/seat/mo).

Anti-slop benchmark (dislop-bench v2)

Run the full offline benchmark (112 cross-modality detection samples across Node & Python + 18 end-to-end writing and design tasks comparing raw AI, naive humanizers, and dislop-guided revisions):

node scripts/run_benchmark.mjs --assert

Latest measured baseline (benchmarks/REPORT.md and benchmarks/PATTERN-BENCHMARK.md):

  • Core Detection (Suite A, 112 samples = 48 v1 + 16 holdout-v1 + 48 text-hard-v2): Precision 100.0% · Recall 100.0% · Clean FPR 0.0% · Adversarial Near-Miss FPR 0.0% · 2nd-Order Trap Recall 100.0% · Node/Python Rule & Statistical Profile Parity 100.0%.
  • Multi-Angle Pattern Benchmark (154 samples = 48 + 16 + 62 pattern-challenge-v2 + 28 holdout-v2): Precision 100.0% · Recall 100.0% · F1 100.0% · Clean FPR 0.0% · Node/Python Parity 100.0%.
  • Bundled Pocket ML (stealth-vapor-v1, 24 paired samples; 36 blind eval with holdouts): 14.6 KB gzip (49.0 KB raw), < 0.5 ms, 100.0% F1, 0.0% Clean FPR.
  • End-to-End Revision (Suite B, 18 tasks = 12 revision-v1 + 6 revision-hard-v2, weights D1=45 / D2=20 / D3=25 / D4=10): raw_ai 60.8 / 100 → naive_humanizer 70.8 / 100 (doubles 3rd-order slop and drops or invents numbers) → dislop_clean 100.0 / 100 (18/18 all-clear).
  • Live 5-model comparison (benchmarks/MODEL-COMPARISON.md): Across Gemini 3.1/3.5 Flash-Lite, Gemini 3.5/3.8 Flash, and Qwen 3.8 Flash, bare prompts averaged 76.0–84.2 / 100 (2/10–4/10 all-clear on the 10-task revision-v1 live subset), while the dislop re-checked cloud pipeline lifted every model by +15.8 to +22.3 points (97.0–100.0 / 100, 8/10–10/10 all-clear; Gemini 3.1 Flash-Lite reached 98.5 / 100 in 1.2s at ~$0.09 / 1k calls).

See benchmarks/README.md to score your own model, Skill, or cloud endpoint (--eval) or export randomized blind human A/B review sheets (--export-blind-sheet).

dislop checks dislop

GitHub Actions CI runs dislop pack-test (164 fixture assertions in both Node and Python), 26 Python unittest conformance and engine tests across 4 modules (test_conformance, test_engine, test_pattern_library, test_pocket_and_local_models), pattern catalog checks (node scripts/pattern_lab.mjs verify and assert), and site checks (tests/test_site.mjs). It also runs dislop check . --fail-on warning across this repository's docs, awesome/, and site/.

Docs & layout

dislop/
├── .agents/plugins/   # OpenAI Codex marketplace manifest (marketplace.json)
├── .claude-plugin/    # Claude Code plugin & marketplace manifests (Skill + local MCP)
├── .codex-plugin/     # OpenAI Codex native plugin manifest & UI card metadata
├── plugin.json        # Agent Plugins 1.0 portable manifest (paired with root mcp.json)
├── gemini-extension.json # Gemini CLI Extension Gallery manifest
├── server.json        # Official MCP Registry manifest (local stdio + remote HTTP)
├── action.yml         # Reusable GitHub Marketplace incremental check Action
├── awesome/           # Awesome Anti-Slop (README.md & README.zh-CN.md, 87 entries)
├── benchmarks/        # Offline anti-slop benchmark datasets (CC0-1.0) & reproducible reports
├── bin/               # Node.js CLI (dislop.mjs) & local stdio MCP launcher (dislop-mcp.mjs)
├── src/
│   ├── core.mjs       # Zero-dep ES module shared by Node CLI, MCP, and browser workbench
│   ├── pattern-engine.mjs # Multi-angle 11-probe extended pattern-bank scanner
│   ├── pocket-model.mjs# Bundled 14.6 KB gzip zero-dep Pocket ML ensemble & domain pack router
│   ├── local-model.mjs# Hardware-aware optional local small-model advisor (Node.js)
│   ├── binary-node.mjs# Zero-dep ZIP reader for Node.js (.docx / .pptx)
│   └── dislop/        # Zero-dep Python 3.10+ engine, pattern scanner, local-model advisor, CLI, and MCP
├── packs/             # Declarative dislop.pack/2 manifests (7 built-in packs, 42 rules, 160 fixtures)
├── patterns/          # Structured 53-pattern anti-slop catalog across 6 dimensions & 7 angles
├── skills/dislop/     # Cross-host Agent Skill (SKILL.md + local ./references/ + openai.yaml)
├── tests/             # 26 Python unittest suites, Node/Python conformance, and site tests
├── scripts/           # Manifest lockstep sync (sync_manifests.mjs), site compiler & benchmarks
├── site/              # Compiled dislop.com workbench, /awesome/, and /llms.txt pages
├── assets/            # Logo mark, light/dark banners, and 1200x630 social preview card
└── docs/              # Brand guide, architecture, design system, roadmap, and research

License

Apache License 2.0

Metadata

Release files for dislop 0.6.2

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for dislop 0.6.2
File Size Uploaded Provenance Actions
dislop-0.6.2.tar.gz 209.3 kB With provenance Details for dislop-0.6.2.tar.gz

Built distribution (wheel)

Table of built distributions (wheels) for dislop 0.6.2
File Interpreter ABI Platform Actions
dislop-0.6.2-py3-none-any.whl Python 3 none any Details for dislop-0.6.2-py3-none-any.whl

Total release size: 402.2 kB

Release files / dislop-0.6.2.tar.gz

Download URL dislop-0.6.2.tar.gz
Size 209.3 kB
Tags Source
SHA-256 checksum
How to use checksums
0891d773103765151bfb079ff7f94930283d207afea345b9c6504490193aea93
BLAKE2b-256 checksum
How to use checksums
e34c364ccfc39b306429b03c6c095067872eccefea4acbd12dc8d3b20fe924f9
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.16

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 22, 2026.

Transparency log

Release files / dislop-0.6.2-py3-none-any.whl

Download URL dislop-0.6.2-py3-none-any.whl
Size 192.9 kB
Tags Python 3
SHA-256 checksum
How to use checksums
05b3e2618a847501f8d2a9efa6e06413e179033671dfbd26c2baf2cc84dcdf2d
BLAKE2b-256 checksum
How to use checksums
808ca5d723576e2e5af21260bb041d968b1f483093ac896048849e92dde13d30
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.16

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 22, 2026.

Transparency log

Security for dislop 0.6.2

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

Full provenance

Both files in this release have been uploaded with provenance metadata.

Select from 2 files

Attestation for dislop-0.6.2.tar.gz

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 22, 2026.

Transparency log

A note on package security

PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.

To learn more about protecting yourself from malicious actors, visit our package security documentation.

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page