Aperture

clfhaha1234/auto-itera

Use when the user wants to autonomously search for the best AI/engineering approach across competing candidates (prompts, models, retrieval strategies, architectures, algorithms) — give it a goal + candidate arms + success threshold, it runs the experiment to a defensible ship-or-kill verdict. Autonomously handles sourcing real production data, scoring arms in parallel, diagnosing per-row, sprint-and-generalize iteration, and writing the conclusion doc. Built-in safeguards (held-out test discipline, variance-floor checks, generalization gates) keep the verdict trustworthy.

tuwulalo/ai-slop-cleaner-en-ru

Detect and clean AI-generated text ("AI slop"), bilingual EN + RU. Use to rewrite AI clichés, "it's not just X, it's Y", "delve", "under the hood", "важно отметить", "под капотом", the rule of three, into natural human prose, or to check whether text was written by a neural network / ChatGPT and rate its "AI-likeness". Detects the text's register first (official / press / blog / forum) and rewrites within it: no slang in a formal post, no officialese in a casual one. Tells decorative AI emoji (🚀💡✅ markers) from genuine human emoji. Двуязычный детектор и чистильщик ИИ-текста: проверить, напис

tangwenwen-md/chinese-de-ai-writing

中文去 AI 味：审稿、改稿、写作、流水线审核四种模式，专管简体中文。规则按证据分级：只有人机语料实测成立的特征才当硬规则，社区共识特征要"命中且空泛"才改，已被实测推翻的流行规则（调句长、删设问、删比喻、加口语词、加错别字、同义词替换）明确不用。 最高约束是信息守恒：不编造经历、病例、数字、引文，不改论断强度。医学内容尤其如此。 附 scripts/zh_ai_scan.py：量化扫描破折号、翻案腔、提示性冒号、段首零回指等 10 项实测特征，给出每千字密度和人机基线对照。 文体分档：小红书/社媒、口播稿与短剧台词旁白、科普、正式文书（标书、汇报、公文）、中文学术。 触发词：「去AI味」「去 AI 味」「AI味太重」「AI腔」「一眼AI」「像AI写的」「不像人话」「说人话」「改得像人写的」「这段有没有AI味」「帮我审一下AI味」「口播稿太书面」「台词不像人说的」「小红书文案太AI」「标书AI味」「中文humanize」。 也用于 Claude 自己产出中文正文（小红书文案、科普稿、台词、汇报）之后的自检。 不管：英文文本；以"降低知网/朱雀 AIGC 检测率"为唯一目标的改写（本 skill 改的是写法，不承诺过检测，也不用错别字、回译等对抗手段）；视觉层面的 AI 味；剧情结构与留存。

madrank8/ai-citation-auditor

Audit any page to predict whether AI search engines (Google AI Overviews and AI Mode, ChatGPT, Perplexity, Gemini, Claude) will cite it, then return a prioritized fix list. Scores 6 dependency-free layers gated by a retrieval-rank prerequisite: answer-first extractability, evidence corroboration, consensus alignment, source-quality fingerprint, SEO-slop detection, and machine structure. Runs on native web fetch and search or on pasted content, with no paid APIs, connectors, or crawlers required. Two modes: AUDIT (URL or draft in, scored scorecard plus fixes out) and REWRITE (turn one section i

cassmtnr/claude-code-starter

Analyze a project's tech stack and generate comprehensive Claude Code configuration files (.claude/ directory with CLAUDE.md, skills, agents, rules, and commands). Use when setting up Claude Code for a new or existing repository.

devravik/engineering-intelligence

Senior engineering quality control for Antigravity agents (detect, attribute, prioritize, repair, verify).