Remove signs of AI-generated code. Like humanizer for text, but for code.
A Claude Code skill that detects and fixes patterns that make code obviously machine-written: over-engineering, copy-paste proliferation, phantom edge cases, comment noise, test smells, security anti-patterns, and more.
Based on 40+ published research papers and industry reports (2022-2026), including 15+ from 2026. Every pattern includes a citation to real research.
When you run /deai-code on your code, it:
- Scans for 26 research-backed AI code patterns
- Rewrites problematic sections like a senior dev would
- Does a self-audit pass ("What still looks AI-generated?")
- Reports which patterns were found with links to the research
| # | Pattern | Research Source | Key Finding |
|---|---|---|---|
| 1 | Over-engineering | CMU (arXiv:2511.04427) | +25.1% cyclomatic complexity |
| 2 | Refactoring avoidance | GitClear (211M lines, 2025) | Refactoring dropped from 24% to 9.5% |
| 3 | Boilerplate inflation | Faros AI (10K devs, 2025) | +154% PR size |
| 4 | Copy-paste proliferation | GitClear (2024-25) | 4x growth in code clones |
| 5 | Repetitive templates | Liu et al. (arXiv:2504.12608) | 20 repetition patterns identified |
| 6 | Comment noise | OX Security (300+ repos, 2025) | 90-100% prevalence |
| 7 | Phantom edge cases | OX Security (2025) | 90-100% prevalence |
| 8 | Error handling theater | CodeRabbit (470 PRs, 2025) | 2x more error handling gaps |
| 9 | Hardcoded secrets | Truffle Security (2025) | Most LLMs recommend hardcoding |
| 10 | Insecure defaults | NYU (IEEE S&P 2022) | 40% of programs vulnerable |
| 11 | Hallucinated dependencies | Spracklen et al. (USENIX 2025) | 19.7% of packages don't exist |
| 12 | Test magic numbers | Ouedraogo et al. (arXiv:2410.10628) | 20,505 test suites analyzed |
| 13 | Shallow test coverage | OX Security (2025) | 40-70% prevalence |
| 14 | Uniformly long lines | Yang et al. (ACM MSR 2024) | F1=0.91 detection on line length |
| 15 | Generic naming | CodeRabbit (2025) | 2x more naming inconsistencies |
| 16 | Code churn | GitClear (2025) | r=0.98 correlation with AI usage |
| 17 | Vanilla reimplementation | OX Security (2025) | 40-70% prevalence |
| 18 | By-the-book fixation | OX Security (2025) | 80-90% prevalence |
| 19 | Unused constructs | Cotroneo et al. (ISSRE 2025) | 500K+ samples analyzed |
| 20 | Performance-blind code | CodeRabbit (2025) | 8x more perf issues |
| 21 | Smell amplification | Ghosh Paul et al. (arXiv:2510.03029) | +63% more code smells |
| 22 | Bug deja vu | OX Security + Tambon et al. (2024-25) | 80-90% prevalence |
| 23 | Silent failures | IEEE Spectrum (Jan 2026) | Tasks 7-8h vs 5h, code removes safety checks |
| 24 | Readability degradation | Horikawa et al. (NAIST, arXiv:2603.13723, Mar 2026) | Maintainability decreased in 56.1% of commits |
| 25 | Vibe coding debt | ICSE 2026 + ETH Zurich CHI 2026 | 101 sources, speed-quality paradox |
| 26 | Build system smells | Ghammam & Almukhtar (arXiv:2601.16839, Jan 2026) | 364 build smells identified |
# Clone into your Claude Code skills directory
git clone https://github.com/golovatskygroup/deai-code.git ~/.claude/skills/deai-codeThat's it. Claude Code will automatically pick up the skill on next launch.
Start Claude Code and type /deai-code. You should see the skill activate.
Alternatively, check it appears in your skill list:
ls ~/.claude/skills/deai-code/SKILL.mdcd ~/.claude/skills/deai-code && git pull/deai-code
Then paste or reference the code you want to clean up. The skill will:
- Identify AI patterns with research citations
- Provide a draft rewrite
- Self-audit for remaining tells
- Deliver a final clean version
/deai-code review src/auth.ts
Before (AI-generated):
class UserValidationStrategy(ABC):
@abstractmethod
def validate(self, user: User) -> ValidationResult:
pass
class EmailValidationStrategy(UserValidationStrategy):
def validate(self, user: User) -> ValidationResult:
if not re.match(r'^[\w\.-]+@[\w\.-]+\.\w+$', user.email):
return ValidationResult(success=False, error="Invalid email")
return ValidationResult(success=True)
class UserValidator:
def __init__(self, strategies: list[UserValidationStrategy]):
self._strategies = strategies
def validate(self, user: User) -> list[ValidationResult]:
return [s.validate(user) for s in self._strategies]
validator = UserValidator([EmailValidationStrategy()])After (de-AI'd):
def is_valid_email(email: str) -> bool:
return bool(re.match(r'^[\w\.-]+@[\w\.-]+\.\w+$', email))Pattern found: #1 Over-engineering (CMU arXiv:2511.04427 — +25.1% cyclomatic complexity in AI code). Strategy pattern with single implementation, abstract base class never extended, wrapper class adding no value.
Full citations are in SKILL.md. 38 sources total. Key sources by year:
- IEEE Spectrum — "AI Coding Degrades: Silent Failures Emerge" (Jan 2026) — quality plateau and decline
- DryRun Security (March 2026) — 87% of AI agent PRs contained vulnerabilities
- OX Security — 2026 AppSec Benchmark — critical findings quadrupled YoY
- Opsera — "AI Coding Impact 2026 Benchmark" (250K+ devs) — 15-18% more security vulns
- Endor Labs / CMU / Columbia / JHU (March 2026) — only 10% of AI code functional AND secure
- NAIST — Horikawa et al., "Do AI Agents Really Improve Code Readability?" (arXiv:2603.13723)
- ICSE 2026 — "Vibe Coding in Practice" (101 sources, speed-quality paradox)
- ETH Zurich / CHI 2026 — vibe coding proficiency study (100 students)
- SonarSource — 2026 Developer Survey — 96% don't trust AI, only 48% verify
- METR (Feb 2026) — experienced devs still 19% slower with AI
- Veracode — 2026 State of Software Security — 82% of orgs have security debt
- Black Duck — 2026 OSSRA — vulnerabilities per codebase doubled (+107%)
- ICSE/LLM4Code 2026 — Tihanyi et al., 95.8% accuracy on LLM code attribution
- Stack Overflow — "A New Worst Coder Has Entered the Chat" (Jan 2026)
- QCon/Thoughtworks (March 2026) — AI coding agent security incidents now weekly
- GitClear — "AI Copilot Code Quality" reports (2024, 2025) — 211M lines analyzed
- CMU — "Speed at the Cost of Quality: How Cursor AI Increases Code Complexity" (2025)
- Stanford — Perry et al., "Do Users Write More Insecure Code with AI Assistants?" (ACM CCS 2023)
- NYU — Pearce et al., "Asleep at the Keyboard?" (IEEE S&P 2022, Distinguished Paper)
- OX Security — "Army of Juniors" (2025) — 10 anti-patterns across 300+ repos
- CodeRabbit — "State of AI vs Human Code Generation Report" (2025) — 470 PRs analyzed
- USENIX — Spracklen et al., "We Have a Package for You!" (2025) — package hallucinations
- Truffle Security — "LLMs are Teaching Developers to Hardcode API Keys" (2025)
- University of Waterloo — Yang et al., "Whodunit" (ACM MSR 2024) — stylometric detection
MIT