Research & Data

Google の Deep Review（旧 Paper Assistant Tool, PAT）を模した、論文・技術文書の深い検証レビュー。文書を論理セグメントに分け、情報密度に応じて検証の深さ（Light/Medium/High）を配分し、セグメントごとに観点の異なる検証サブエージェントを並列に走らせる（証明の逐行検証・反例探索・数値の再計算・実験設計・主張と証拠の対応・新規性の文献確認）。候補は逐語引用で照合・重複排除し、反証検証で誤検出を落としたうえで、Overall Feedback と Detailed Feedback（Major concerns / Minor corrections / Suggestions）の Markdown・HTML レポートにまとめる。スコアは付けない。「Deep Reviewして」「ディープレビュー」「PATみたいに」「論文の誤りや証明の穴を徹底的に探して」「投稿前に技術的なミスを洗い出して」「数式・証明・実験を検証して」などと頼まれたら、スキル名が出ていなくても使うこと。設計書・技術報告など論文以外の技術文書にも使える。採点つきの会議形式の査読や、査読→改訂のループは paper-review-loop を使う。Use for deep, agentic error-finding review of a paper or technical document in the style of Google's Deep Review / PAT.

論文PDFを、その分野に精通した厳格かつ建設的な査読者として査読し（Paper Summary / Strengths / Weaknesses / Justification / Rebuttal向け質問 / Rating / Confidence）、指摘に沿って原稿を改訂→PDF再生成→再査読するサイクルを、ユーザーが指定した制限時間内でできるだけ多く繰り返す。起動直後に制限時間をユーザーに確認する。論文・原稿・preprint・投稿前やリジェクト後の論文PDFについて「査読して」「レビューして」「採択されるように改善して」「rebuttalに備えたい」「self-review」「ブラッシュアップして」などと頼まれたら、スキル名が出ていなくても使うこと。Federated Learning（non-IID・partial participation・privacy）、機械学習の理論・大規模実験の論文向けの詳しいチェックリストを持つ。Use for peer review of a paper PDF and time-boxed review→revise→re-review loops.

Use when stuck, looping, or tunnel-visioning on a hard problem — forms a committee of two contrasting paseo agents that analyse it in parallel, argue out their disagreement, and converge on a plan. Triggers on \"/pcommittee\", \"committee\", \"we are going in circles\", \"root cause still unclear after several attempts\", \"have two models look at this\". Analysis only; the committee never edits files.

Full-site crawling, scraping, and site mapping via Firecrawl MCP. Use when user says "crawl site", "map site", "full crawl", "find all pages", "broken links", "site structure", "discover pages", "JS rendering", or needs site-wide analysis.

Ahrefs API analyst (extension). Reads referring domains, backlinks, organic keywords, and content explorer data via the tested @ahrefs/[email protected] server. Pairs with seo-backlinks for multi-source confidence weighting.

Search performance analysis, traffic-drop diagnosis and SEO reporting. Use for "why did our traffic drop", "organic traffic fell", "lost rankings", "analyse my Search Console export", "find quick wins in GSC", "striking distance keywords", "low CTR pages", "keyword cannibalisation", "compare this month vs last", "branded vs non-branded", "did the Google update hit us", "analyse server logs / Googlebot crawl", "build an SEO dashboard / KPI report", "did my SEO change work", "rank tracking". Uses gsc_analyze.py, log_analyze.py and Looker Studio (free); explains when Semrush/Ahrefs rank tracking is worth paying for. Not for crawling the site for defects (seo-technical / seo-audit), keyword discovery for new content (seo-keywords), or competitor research (seo-competitors).

Competitor and SERP analysis. Use for "who are my SEO competitors", "why does X outrank me", "compare my page with the top results", "competitor content gap", "keyword gap vs competitors", "what is ranking for <keyword>", "SERP features for <keyword>", "competitor backlinks", "how much traffic does competitor get", "what did competitor change on their page", "monitor competitor sites". Finds real SERP competitors (not just business rivals), compares pages with compare_pages.py, inventories SERP features and intent, tracks competitor page changes via the Wayback Machine, and states exactly when Semrush/Ahrefs/Similarweb are worth paying for. Not for keyword discovery from seeds (seo-keywords), link building execution (seo-links), or your own traffic analysis (seo-analytics).

Use for visibility in AI answers (AEO, GEO, "LLM SEO") — "get cited by ChatGPT / Perplexity / Google AI Overviews / AI Mode / Copilot / Claude / Gemini", "why doesn't ChatGPT mention us", "should we block GPTBot or ClaudeBot", "AI crawler access", "do we need llms.txt", "track our AI visibility", "brand in AI answers", "opt out of AI Overviews". Covers eligibility, crawler policy by purpose (search vs user-fetch vs training), Bing indexing for ChatGPT and Copilot, consistent entity facts, citable content, off-site sources, and repeatable measurement with confidence intervals. Not for: classic keyword and ranking work (seo-keywords, seo-content), schema validation (seo-schema), Business Profile and maps (seo-local), traffic-drop forensics (seo-analytics).

In-place upgrade of legacy PHP 5.x/7.x codebases (custom CMS, TemplatePower, mysql_*, ereg, short tags, PHP4 constructors) to run on PHP 8.3/8.4, then deploy without Docker via a .env config and an FTP/rsync deploy.sh. Use for modernizing legacy PHP CMS/sites and sibling legacy projects, or when a legacy site must be moved to shared hosting (cPanel/Plesk) over FTP.

Use when comparing a draft against specific prior work head to head — checking whether a baseline was trained at the same scale, whether protocols match, how many seeds each ran, or what a competitor actually reports. Triggers on "compare this paper to X", "how does our method differ from Y", "is the baseline comparison fair", "did they train at the same scale", or a shortlist of papers from a gap sweep.

Spend the fewest tokens, tool calls, and words that still give a correct, complete, verified result. Use for every task with more than one step, even when the user does not mention cost, including code changes, debugging, refactoring, repository searches, multi-file edits, research, CI and build work, and any tool-assisted workflow, and whenever the user asks for concise, quick, cheap, or token-efficient work. Not for a one-step answer that needs no tools, such as arithmetic, a greeting, or a short fact.

Select and use shared Dumplings PackageModule, Core, PowerHTML, and powershell-yaml functions for networking, redirects, response decoding, temporary files, archive extraction, text and structured data, time conversion, update feeds, Playwright browser access, HTML parsing, and YAML processing. Use alongside installer analysis, WinGet manifest authoring, or Dumplings task authoring whenever commands or scripts need project helper APIs.

Use when a request starts from an industry, market, or value chain and asks where revenue and profit pools sit, how representative business models compare, why economics changed, or which external startup entry points are credible. Also use for sector-wide comparisons of margins, capital intensity, cash conversion, and control points. Do not use for a company-first two-layer review.

Selects, calculates and explains KPIs for any tabular dataset (CSV/TSV) in any field. Use when the user attaches data and asks for KPIs, metrics, performance indicators, efficiency, error metrics such as RMSE or MAE, financial ratios, or to evaluate predictions against actuals, optionally naming KPIs to include or exclude, a research focus, or papers and earlier results whose KPIs should be applied to the data.

Build, extend, or audit a reproducible analysis pipeline of nested bash scripts in a project's pipeline/ directory, so that one command (bash pipeline/run_pipeline.sh) recreates every output from raw data. Use when the user wants to set up a pipeline for an R/Quarto research project, add a new script or stage to an existing pipeline/, reorder scripts by dependency, check that every analysis is covered by the pipeline, or make a project reproducible end to end.

Pre-submission review of your own CS paper draft, with extra checks for security and privacy papers. Use when asked to "review my paper/draft", "pre-submission review", "critique this paper", "is my paper ready for CCS" or "what would reviewers say". Runs the review in a fresh subagent and saves it as Markdown.

Turns a feature idea into a one-page spec covering problem, audience, MVP scope, non-goals, numbered Given/When/Then acceptance criteria, affected routes, data and emails, and open questions. Use when the user says "brief", "spec this", "scope this" or "what would it take", or before starting a feature whose scope is unclear.

Distill a bloated answer or messy topic into concise Discussion and Analysis prose. Use when an answer ran long, or when the user calls onepager (/onepager in Claude Code, $onepager in Codex).

多视角交叉验证深度调研——从外部真实结构读出多个正交视角，各视角独立并行检索取证，按显式预算收敛，产出带"证据收敛等级"（收敛/确证/单源/争议/未找到）的内部底稿，再由一个独立 pass 润色成机制不可见、附完整来源与交叉验证的读者稿 reader-report.md；该文件全文原样作为对用户的最终回复。触发条件（命中即用，无需点名本 skill）：用户说"深度调研 / 深度研究 / 深度分析 / 深入研究 / deep-research / deep research"，或提出同义要求——仔细调研、认真调研、调研得仔细一点、多查几个来源、确定性调研、要确定性结论、别只给一个来源、多维度调研、多视角、全景分析、正反双方梳理、交叉验证、多方观点对比、彻底搞清楚 X、把 X 查透、X 到底怎么样，或要求标注"哪些结论有多来源印证、哪些只有单一来源"。与 anysearch / agent-reach 的通用"调研/搜索"触发重叠时，凡带"深/仔细/确定性/交叉验证"限定词者以本 skill 为准。不适用于一次性快速查证（那用 anysearch 或 agent-reach 即可）。

Deep legal research that fans out multiple agents across free legal sources, verifies every authority, and returns a polished, well-cited memo. Use for case-law research, legal-scholarship / literature reviews, or both. Modes — caselaw | scholarship | both. Trigger when the user wants a thorough doctrinal memo, a survey of the academic debate, an answer grounded in cases/statutes, or a research deep-dive on a legal question.

Quick, verified citation-finding for a specific proposition — a footnote-filler, not a research memo. Returns a few Supreme Court cites and/or a few scholarly cites (top scholars in flagship journals preferred; student notes only as a last resort), each verified with a pincite and a parenthetical or quote. Trigger when the user asks "find me a cite for X", "what should I cite for this footnote", "I need authority for the proposition that…", or pastes a sentence with an empty footnote.

Build or extend a Zotero collection of verified sources for a project — find the sources (web, catalogs, HeinOnline via an institutional library proxy), add them with full metadata via the Zotero Web API, then run a deterministic anti-hallucination check that every added item must pass. Trigger when the user wants sources gathered into Zotero ("build a library for X", "add the books/articles on X to my Zotero", "populate <collection> with …").

>- Use when connecting Croct to a third-party platform or migrating to Croct from another personalization or experimentation tool. Covers Storyblok (Optimize by Croct app, @croct/plug-storyblok, Storyblok withCroct, resolveContent, croct/plug-storyblok, CroctStoriesApi), Sanity and Strapi (slot fallback, dynamic zones, metadata.contentSource), Shopify Hydrogen storefront events, Shopify custom pixel (croctPixel, PUBLIC_CROCT_SHOPIFY_PIXEL, checkout and order tracking), VTEX, Google Tag Manager data layer, Microsoft Clarity tags, HubSpot forms (hbspt.forms.create, hs-form-event), firmographic providers 6sense, Demandbase, Clearbit Reveal, and ZoomInfo WebSights (firmographics attributes, account-based audiences in CQL), and migrations from Mutiny and Ninetailed (Experience component, Contentful sunset). Triggers include cdn.croct.io/plug-integration scripts, "personalize Storyblok blocks", "track Shopify checkout", "send VTEX events", "ABM personalization", "replace Mutiny".

Use this skill when the user's prompt requires (1) researching a topic across multiple sources, comparing options or alternatives, analyzing trends or history, understanding markets or industries, or reviewing literature or studies and (2) synthesizing that research into a comprehensive, narrative report. If you're planning to search the web or internal knowledge bases, consider using this skill. This skill coordinates research subagents, so use it only when you have a tool for spawning subagents (the Agent or Task tool); otherwise, research the question directly.