0xwilliamortiz
Featured
ponytail-improved
Ponytail is a plugin for AI coding agents that enforces a disciplined ladder of reuse before writing code, reducing code volume by roughly 54% while preserving safety.
anti-slop
Anti-slop is a tool for detecting and repairing substance defects in AI-assisted prose, code, documentation, and agent output using deterministic scanners and structural procedures that produce verifiable artifacts rather than authorship judgments.
Stars
6
7-day growth
No data
Forks
2
Open issues
0
License
Apache-2.0
Last updated
2026-07-28
Why it is worth attention
It rejects the broken approaches of AI detectors (biased, unreliable) and humanizers (degrade quality), offering a principled, evidence-based alternative that never emits authorship verdicts and is backed by a rigorous source ledger with 43 entries and eight corrections to commonly repeated figures.
Who it is for
- Writers and editors creating AI-assisted content who want to ensure substance over style
- Developers and teams integrating AI-generated code or documentation and need to catch structural defects
- Researchers studying AI writing quality, detection methodology, and citation integrity
- Claude Code users seeking to enforce content quality standards in their development workflow
Use cases
- Running deterministic scanners on markdown files to find placeholder text, unresolved references, or dependency issues
- Using the Claude Code plugin for slop review and rewrite during iterative development
- Verifying citations and claims in AI-assisted research or technical documentation
- Checking codebases for AI-generated structural problems like wrapper tests that do not fail
Strengths
- Evidence-backed: every numeric claim traces to a source ledger with 30 primary sources, each with URL, retrieval date, evidence tier, and limitations
- Transparent: five structural tests produce artifacts (cut spans, negations, resolved citations) that anyone can inspect and verify
- Principled design: never emits authorship verdicts, never hard-fails on stylistic markers alone, and requires a procedure to convict
- Comprehensive toolset: includes deterministic scanners, evidence-tiered markers, a Claude Code plugin with five skills, and a 62-file knowledge base with 314 tests
Considerations
- Does not detect AI authorship or score writing quality holistically by design, limiting its use for attribution or general quality assessment
- Marker lists are incomplete and non-durable; vocabulary shifts by model generation and human speech converges on model vocabulary
- The AI-code-quality literature is not settled; the strongest pre-registered study found no significant differences, while observational studies found opposite effects
- The tool's own notes may contain the defects it describes, and while self-scanned in CI, absolute freedom from defects is not claimed
README quick start
anti-slop
Find and repair substance defects in AI-assisted prose, code, documentation, and agent output.
It reports defects. It never reports authorship.
Why this is not another AI detector or humanizer
Most tools in this space do one of two things, and both are broken.
Detectors guess who wrote something. They are unreliable and their failures land on identifiable people. Sixteen detection models disproportionately flagged English-language-learner essays, and non-White ELL students more than White ELL peers, while human annotators on the same essays showed no significant demographic bias (Stowe et al., ACL 2026). OpenAI withdrew its own classifier at 26 percent true positive and 9 percent false positive.
Humanizers strip the surface tells. The Wikipedia guide that most of this field derives from warns against exactly that, in bold:
The patterns listed here are also only potential signs of a problem, not the problem itself. Please do not merely treat these signs as the problems to be fixed; that could just make detection harder.
The measurement agrees. "All humanizers tend to degrade the quality of the original text": best-tier tools win a fluency comparison against the original only 26.0 percent of the time (DAMAGE, COLING 2025).
And you cannot just ask a model. Agreement between LLM judges and human slop labels is kappa 0.01 for GPT-5, minus 0.01 for DeepSeek-V3, and 0.03 for o3-mini, which is chance (Shaib et al., arXiv 2509.19163). Worse, judges are biased toward slop features: GPT-4 preferred model-written pitches 89 percent of the time against human raters at 36 percent (PNAS 122(31)).
So this project never asks whether text "looks like" slop. Every check is a mechanical procedure that emits an artifact you can inspect.
How it works
Three layers, and only the bottom two may fail a build.
%%{init: {"theme":"base","themeVariables":{"fontFamily":"ui-sans-serif, -apple-system, Segoe UI, Helvetica, Arial, sans-serif","fontSize":"14px","lineColor":"#A6A29C","textColor":"#3A3A3C","edgeLabelBackground":"#FFFFFF","secondaryTextColor":"#3A3A3C","tertiaryTextColor":"#3A3A3C","background":"#F7F6F3"},"flowchart":{"curve":"basis","nodeSpacing":45,"rankSpacing":60}}}%%
flowchart LR
IN["Draft, diff, docs,agent output"]
L0["Layer 0deterministic scanners"]
L2["Laye
Description
Find and repair substance defects in AI-assisted prose, code, docs, and agent output. Reports defects, never authorship. Structural tests over model judgement, because LLM judges agree with human slop labels at chance.
Related repositories
Similar projects matched by category, topics, and programming language.
makecindy
Featured
cindy
Cindy is an open-source AI agent that runs locally on your machine, integrates multiple AI harnesses and models, and provides memory, skills, and automation to perform real work in your projects and apps.
0xwilliamortiz
Featured
openclaude-improved
OpenClaude is an open-source CLI coding agent that runs on any platform and supports a wide range of LLM providers, offering the same tools and workflows as Claude Code.