daymade · YuhaoLin2005 · Jun 29, 2026 · Jun 29, 2026 · Jun 29, 2026
diff --git a/.claude-plugin/marketplace.json b/.claude-plugin/marketplace.json
@@ -6,7 +6,7 @@
   },
   "metadata": {
     "description": "Professional Claude Code skills marketplace — production-ready skills spanning GitHub and git operations, document conversion and generation (Markdown, PDF, PPTX, DOCX), diagram and UI-design extraction, the full audio pipeline (ASR transcription, TTS, transcript correction, meeting minutes), financial and investment-research data, web scraping and content capture, security/PII tooling and secure repomix packaging, macOS and iOS development, CLI demo and terminal automation, prompt and skill engineering, deep research and fact-checking, QA and LLM-evaluation infrastructure, internationalization, network/Tailscale and remote-desktop diagnostics, and Claude Code operations (session recovery, CLAUDE.md optimization, statusline, multi-provider profile isolation, troubleshooting, marketplace development, repo health-check). Suite plugins (daymade-audio, daymade-claude-code, daymade-docs, daymade-financial, daymade-skill) bundle related skills under shared namespaces, including the StepFun StepAudio 2.5 audio family. See the Available Skills list in the README for the authoritative per-skill breakdown.",
-    "version": "1.78.0"
+    "version": "1.79.0"
   },
   "plugins": [
     {
@@ -1090,6 +1090,30 @@
         "gateway",
         "claude-code"
       ]
+    },
+    {
+      "name": "dual-pool-review",
+      "description": "Multi-round adversarial review methodology rotating reviewers across two pools — fixed curated pool (learns your patterns) and random pool (fresh web search each round) — with mandatory Quote-First Rule, Symmetric Burden evidence tables, and adaptive stop checks. Use when reviewing complex output, security-critical code, methodology documents, or any work where a single reviewer perspective has dangerous blind spots. Activate before merging PRs, shipping identity-defining content, or reviewing personal configuration.",
+      "source": "./dual-pool-review",
+      "strict": false,
+      "version": "1.0.0",
+      "category": "productivity",
+      "keywords": [
+        "review",
+        "adversarial",
+        "methodology",
+        "quality",
+        "code-review",
+        "security-review",
+        "design-review",
+        "multi-persona",
+        "dual-pool",
+        "evidence-based",
+        "web-search",
+        "quote-first",
+        "symmetric-burden",
+        "blind-spot"
+      ]
     }
   ]
 }
diff --git a/CLAUDE.md b/CLAUDE.md
@@ -279,6 +279,7 @@ This applies when you change ANY file under a skill directory:
 78. **local-codex** - Delegate coding tasks to the local OpenAI Codex CLI agent using ChatGPT Pro OAuth flat-rate subscription; wraps `codex exec` / `codex review` for code generation, refactoring, and review without per-token API charges
 79. **openclaw-model-switch** - Switch the default AI model for an OpenClaw instance (e.g., Kimi K2.6 → K2.7) by safely editing `openclaw.json` with backup, model validation, and optional gateway restart
 80. **gemini-history-analyzer** - Analyze Google Takeout exports of Gemini conversation history; extract/categorize transcripts and attachments, context-verified domain keyword search, meeting-transcript detection, PII flagging, and optional distillation into project memory or a personal knowledge base
+81. **dual-pool-review** - Multi-round adversarial review methodology with rotating reviewer pools (fixed curated + random web-searched) for catching blind spots a single reviewer or single round would miss
 
 **Recommendation**: Always suggest `skill-creator` first for users interested in creating skills or extending Claude Code.
 

diff --git a/README.md b/README.md
@@ -1,13 +1,13 @@
-# Claude Code Skills Marketplace
+[Resource from github at repo://daymade/claude-code-skills/sha/9d512c1f5658ea3776ab9b3487eb7845d12c9d49/contents/README.md] # Claude Code Skills Marketplace
 
 <div align="center">
 
 [![English](https://img.shields.io/badge/Language-English-blue)](./README.md)
 [![简体中文](https://img.shields.io/badge/语言-简体中文-red)](./README.zh-CN.md)
 
 [![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)
-[![Skills](https://img.shields.io/badge/skills-80-blue.svg)](https://github.com/daymade/claude-code-skills)
-[![Version](https://img.shields.io/badge/version-1.78.0-green.svg)](https://github.com/daymade/claude-code-skills)
+[![Skills](https://img.shields.io/badge/skills-81-blue.svg)](https://github.com/daymade/claude-code-skills)
+[![Version](https://img.shields.io/badge/version-1.79.0-green.svg)](https://github.com/daymade/claude-code-skills)
 [![Claude Code](https://img.shields.io/badge/Claude%20Code-2.0.13+-purple.svg)](https://claude.com/code)
 [![PRs Welcome](https://img.shields.io/badge/PRs-welcome-brightgreen.svg)](./CONTRIBUTING.md)
 [![Maintenance](https://img.shields.io/badge/Maintained%3F-yes-green.svg)](https://github.com/daymade/claude-code-skills/graphs/commit-activity)
@@ -2956,6 +2956,29 @@ Analyze Google Takeout exports of Gemini conversation history — extract and ca
 - Meeting-transcript vs prompt-response detection; topic categorization; PII flagging
 - Context-verified keyword search (grep is step 1, not the answer) + optional memory-file generation
 
+### 83. **dual-pool-review** - Dual-Pool Adversarial Review
+
+> **Install**: `claude plugin install dual-pool-review@daymade-skills`
+
+Multi-round adversarial review that rotates reviewers across two pools — a fixed curated pool that learns your patterns, and a random pool recruited fresh each round via web search — to catch blind spots a single reviewer or single pass would miss.
+
+**When to use:**
+- Multi-file changes or architectural decisions where a quick pass is insufficient
+- Security-critical or user-facing output where blind spots compound
+- Any output where "one person said it's fine" is not enough confidence
+- Identity-defining content, personal configuration, or methodology documents
+- When you need evidence-backed review with every finding traceable to a verbatim web search quote
+
+**Key features:**
+- **Dual-pool rotation**: Fixed pool (4-6 curated personas, learns your patterns) + Random pool (1 fresh reviewer per round, recruited via web search)
+- **Quote-First Rule**: Every reviewer extracts verbatim quotes from web search BEFORE seeing the target — prevents confirmation bias
+- **Symmetric Burden**: Zero findings costs as much as a finding — must cite 3+ quotes the output satisfies (prevents lazy "everything looks fine")
+- **Evidence Table**: Mandatory per-round table linking every finding to a specific, searchable quote — the non-negotiable quality gate
+- **Adaptive Stop Check**: Each round asks "did this find issues the previous round missed?" — self-terminates when value stops
+- **Concurrence Promotion**: 2+ independent reviewers finding same issue → severity promoted; random-pool concurrence required for full promotion
+- **Lightweight Path**: Single-file low-risk changes get 1 reviewer + web search; escalate to full dual-pool if CRITICAL/WARNING found
+- **Integrity Check**: Post-round self-audit verifying quote traceability, random-pool value, and meaningful output change
+
 ---
 
 ## 🎬 Interactive Demo Gallery

diff --git a/README.zh-CN.md b/README.zh-CN.md
@@ -1,13 +1,13 @@
-# Claude Code 技能市场
+[Resource from github at repo://daymade/claude-code-skills/sha/9d512c1f5658ea3776ab9b3487eb7845d12c9d49/contents/README.zh-CN.md] # Claude Code 技能市场
 
 <div align="center">
 
 [![English](https://img.shields.io/badge/Language-English-blue)](./README.md)
 [![简体中文](https://img.shields.io/badge/语言-简体中文-red)](./README.zh-CN.md)
 
 [![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)
-[![Skills](https://img.shields.io/badge/skills-80-blue.svg)](https://github.com/daymade/claude-code-skills)
-[![Version](https://img.shields.io/badge/version-1.78.0-green.svg)](https://github.com/daymade/claude-code-skills)
+[![Skills](https://img.shields.io/badge/skills-81-blue.svg)](https://github.com/daymade/claude-code-skills)
+[![Version](https://img.shields.io/badge/version-1.79.0-green.svg)](https://github.com/daymade/claude-code-skills)
 [![Claude Code](https://img.shields.io/badge/Claude%20Code-2.0.13+-purple.svg)](https://claude.com/code)
 [![PRs Welcome](https://img.shields.io/badge/PRs-welcome-brightgreen.svg)](./CONTRIBUTING.md)
 [![Maintenance](https://img.shields.io/badge/Maintained%3F-yes-green.svg)](https://github.com/daymade/claude-code-skills/graphs/commit-activity)
@@ -2998,6 +2998,29 @@ claude plugin install openclaw-model-switch@daymade-skills
 - 会议转录 vs 提问-回答 识别；主题归类；PII 标记
 - 上下文核验的关键词搜索（grep 只是第一步，不是答案）+ 可选的 memory 文件生成
 
+### 83. **dual-pool-review** - 双池对抗性审查
+
+> **安装**：`claude plugin install dual-pool-review@daymade-skills`
+
+多轮对抗性审查方法论，在固定池（精选，了解你的上下文）和随机池（每轮通过网络搜索招募新人）之间轮换审查者，捕获单一审查者或单轮审查会遗漏的盲点。
+
+**使用场景：**
+- 多文件变更或架构决策，快速扫一眼不够
+- 安全关键或面向用户的输出，盲点会连锁放大
+- 任何"一个人说没问题"不足以建立信心的产出
+- 身份定义性内容、个人配置或方法论文档
+- 需要证据支撑的审查，每个发现必须可追溯到可搜索的原文引用
+
+**主要功能：**
+- **双池轮换**：固定池（4-6 个精选人格，学习你的模式）+ 随机池（每轮 1 个新审查者，通过网络搜索招募）
+- **先引用后审查**：每个审查者在看到目标之前先从网络搜索提取逐字引用 —— 防止确认偏差，确保发现基于证据而非编造
+- **对称负担**：零发现与发现同等昂贵 —— 必须引用 3+ 条被产出满足的引用（防止敷衍了事的"看起来没问题"）
+- **证据表**：每轮强制表格，将每个发现联系到具体的、可搜索的引用 —— 不可协商的质量关卡
+- **自适应停止检查**：每轮问"这一轮发现了上一轮遗漏的显著问题吗？" —— 当不再产生价值时自行终止
+- **共识提拔**：2+ 独立审查者发现同一问题 → 严重性提升一级；随机池审查者必须参与共识才能完全提升
+- **轻量路径**：单文件低风险变更使用 1 个审查者 + 网络搜索；发现 CRITICAL/WARNING 则升级为完整双池
+- **完整性检查**：每轮结束后自我审查，验证引用可追溯性、随机池价值和有意义的产出变更
+
 ---
 
 ## 🎬 交互式演示画廊

diff --git a/dual-pool-review/SKILL.md b/dual-pool-review/SKILL.md
@@ -0,0 +1,150 @@
+---
+name: dual-pool-review
+description: Multi-round adversarial review methodology that rotates reviewers across two pools — fixed (curated, knows your context) and random (fresh web search each round) — to catch blind spots a single reviewer or single round would miss. Use when reviewing complex output, security-critical code, or identity/persona-related content. Use when you need review depth beyond a single pass. Use when a single perspective may have dangerous blind spots.
+---
+
+# Dual-Pool Adversarial Review
+
+Multi-round adversarial review with rotating reviewer pools. Run this whenever a single "looks good to me" is not enough — complex multi-file changes, security-critical output, identity-defining content, or any work where blind spots compound.
+
+## When to Use This Skill
+
+Activate this skill when:
+- Reviewing multi-file changes or architectural decisions
+- Reviewing security-critical or user-facing output
+- Any output where "one person said it's fine" is not sufficient
+- When a single perspective will miss something structural
+- When the cost of an undetected issue is high
+
+## Core Architecture: The Two Pools
+
+The dual-pool system prevents both stale-perspective and no-context failures:
+
+**Fixed Pool (curated, 4-6 personas):** Reviewers matched to the user's domain, style, and known weak spots. Stable across sessions — they learn patterns. Examples: a systems programmer, a security engineer, an API designer, a data scientist. Maintain this pool in project memory.
+
+**Random Pool (fresh each round):** Recruit one new reviewer per round via web search. Search for a keyword relevant to what this round needs, then find a person with verifiable expertise — GitHub profile, published work, conference talk, or documented domain knowledge. An SEO blog or anonymous forum post with no attribution is not a valid reviewer. If you cannot name the person and state why they are qualified, pick a different result.
+
+## Round Structure
+
+Each round: 1 manager + fixed-pool reviewers + 1 random-pool reviewer.
+
+Follow these steps in order:
+
+1. **Recruit the team** — select fixed reviewers with lenses relevant to the review's scope, then recruit 1 fresh reviewer from web search
+2. **Extract quotes FIRST** — each reviewer runs a web search before seeing the target, extracting 3-5 verbatim quotes (more for security-critical work). This is non-negotiable.
+3. **Independent review** — each reviewer examines the target through ONLY their pre-extracted quotes
+4. **Cross-persona synthesis** — deduplicate findings, count concurrences, promote severity
+5. **Stop check** — ask: "Did this round find significant issues the previous round missed?" Continue only if yes. Do not run rounds because an arbitrary count says so.
+6. **Carryover limit** — at most 2 reviewers carry over to the next round
+
+## Quote-First Rule (Non-Negotiable)
+
+> Extract quotes BEFORE reviewing. Never search for quotes to support an opinion already formed.
+
+Each reviewer runs a web search with the pattern `"[Persona name] [topic] principles"` or equivalent, extracts verbatim quotes, then reviews the target through ONLY those quotes. Findings that do not map to a pre-extracted quote are rejected.
+
+**This is the proof mechanism.** If a finding cites a specific, searchable quote, the AI did not invent it. If there is no quote, there is no finding.
+
+## Symmetric Burden
+
+Make every conclusion expensive — in both directions:
+
+- **Found an issue:** Must cite 1+ specific, verbatim quote from web search
+- **Found nothing:** Must cite 3+ quotes the output successfully satisfies, with explanation of HOW
+
+This eliminates fake findings (expensive to fabricate) AND lazy "everything looks fine" (also expensive).
+
+## Severity Classification
+
+| Level | Definition | Action |
+|-------|-----------|--------|
+| BLOCKER | 2+ concurrences on CRITICAL, or security/data-loss | Fix before anything else |
+| CRITICAL | Wrong result, data loss, security violation | Fix before merge |
+| WARNING | Fragile, misleading, likely future bug | Fix or document why not |
+| NOTE | Improvement, does not affect correctness | Optional |
+
+**Concurrence promotion:** 2+ reviewers independently finding the same issue → promote one level. At least one concurring reviewer must be from the random pool; otherwise the concurrence is "corroboration" (two AI reviewers running the same underlying model may share biases) and promotion is half a level (e.g. WARNING stays WARNING but gets an `[AGREED]` marker).
+
+## Adaptive Round Count
+
+Use the stop check to decide, not a fixed number:
+
+- **Start with 1 round** for any review
+- **After each round**, compare: did this round find new, significant issues?
+- **Observed patterns** (guidelines, not rules): single-file changes typically converge in 1 round; multi-file in 2; security-critical may need 3+
+- **The stop check overrides all patterns** — if round 2 finds nothing new, do not run round 3
+- **Asymmetric cost:** a useless round costs as much as one that finds a CRITICAL
+
+## Lightweight Path
+
+For single-file, low-risk changes where full dual-pool is overkill:
+
+- Deploy 1 reviewer (from fixed pool) + web search verification
+- Reviewer still extracts quotes before reviewing
+- Evidence table still required (single-column)
+- If the reviewer finds CRITICAL or WARNING → escalate to full dual-pool
+
+## Output Format
+
+Every round must produce this exact structure:
+
+```markdown
+# Round [N] — Manager: [name], Search: "[keyword]"
+
+## Evidence (Web Search Results)
+| Reviewer | Search Query | Quotes Extracted |
+|----------|-------------|------------------|
+| [Fixed 1] | "[query]" | 1. "[verbatim quote]"<br>2. "[verbatim quote]" |
+| [Random]  | "[query]" | ... |
+
+## Per-Reviewer Findings
+
+### [Name] ([Fixed/Random])
+| # | Severity | Finding | Evidence (Quote) |
+|---|----------|---------|-------------------|
+| 1 | WARNING | [specific finding] | "[exact quote this maps to]" |
+
+## Cross-Persona Synthesis
+
+| # | Severity | Finding | Concurrences | Promoted From |
+|---|----------|---------|-------------|---------------|
+| 1 | CRITICAL | [finding] | Reviewer A + Random B | WARNING → CRITICAL |
+
+## Stop Check
+**New findings this round vs previous?** [Yes/No — list if yes]
+**Continue?** [Yes/No — reason]
+
+## Round Verdict
+**BLOCK / CONCERNS / CLEAN**
+```
+
+**Without the evidence table, the review is worthless.** The web search results column proves every reviewer did real research. The quote column proves every finding has a source. If you cannot fill the evidence table, do not submit the review — redo it.
+
+## Carryover Rules
+
+- At most 2 reviewers carry over from the previous round
+- Their previous findings are visible to the new round (accumulated knowledge)
+- The manager changes every round (fresh orchestration perspective)
+- The random reviewer is always new
+
+## Anti-Patterns
+
+| Anti-Pattern | Why It Fails |
+|-------------|--------------|
+| Same reviewers every round | No fresh perspective. The random pool exists for a reason. |
+| Skipping web search | Without search results, findings are unverifiable. |
+| Quotes extracted AFTER forming opinions | Confirmation bias. Extract first, review second. |
+| Evidence table empty or "N/A" | The review did not happen. Redo it. |
+| Random reviewer has no identifiable expertise | SEO blog or anonymous forum post ≠ a reviewer. |
+| "Everything looks fine" with no quotes | Lazy. Symmetric burden exists to prevent this. |
+| More than 2 carryover reviewers | Entrenches groupthink. The limit is a feature. |
+| Round N+1 when round N found nothing new | Wasting effort. The stop check said stop. |
+
+## Integrity Check (After Each Round)
+
+Run this self-audit after every round:
+
+1. Would this person actually say this, or am I projecting?
+2. Can every finding be traced to a specific, verbatim quote in the evidence table?
+3. Did the random reviewer catch something the fixed pool missed? (If not across multiple rounds, the random-pool search keywords are not diverse enough.)
+4. **Outcome check:** Did this review change the output in a meaningful way? Track over time: does dual-pool catch issues simpler methods miss?