Home / All Skills / Workflow & Core Tools / devils-advocate
English (EN)简体中文 (ZH)日本語 (JA)
Workflow & Core Tools · daymade/claude-code-skills

devils-advocate

Structurally pressure-test a thesis (devil's advocate): decomposes the thesis into explicit claims and implicit assumptions (numbered A1/A2), systematically retrieves counter-arguments with direct citations from user-provided evidence, assigns High/Medium/Low risk flags, complements with base-rate outside views, and provides a list of "monitorable falsification signals"; delivers a dual-layer output of structured findings + human-readable narrative report. If no counter-evidence is found, honestly states "Not found" without fabrication. Replicated from LinqAlpha's (hedge fund AI product) public Devil's Advocate architecture, layered with CIA Structured Analytic Techniques and RAND Assumption-Based Planning. Use this skill when the user wants to "pressure test / play devil's advocate / find blind spots / break down assumptions / challenge this thesis / build a bear case / red-team this proposal / review before the Investment Committee / pressure test my thesis", or brings an investment thesis, IC memo, research conclusion, or business decision to be critiqued—even if they do not explicitly say "devil's advocate".

L2 Local · Restricted workspaceV1 Commit-PinnedWrite FileDual-pass Audit
Author daymadeGitHub daymade/claude-code-skills ↗Stars 1385License MIT（仓库根 LICENSE，Copyright (c) 2025 daymade；GitHub API spdx MIT）Commit d5c4678cb5
Control levels describe skill capabilities and suggested control measures; they do not represent quality or safety certifications, nor do they dictate usability. Usage decisions rest entirely with users and administrators. Different agent hosts apply varying constraints; do not assume hosts provide automatic protection. Third-party skill copyrights belong to their original authors. Use at your own risk.

1How It Works & Mechanism

Methodologically replicates the production prompt/schema publicly shared by LinqAlpha (hedge fund AI product) for Devil's Advocate, layered with techniques such as CIA Key Assumptions Check, ACH, and RAND ABP; replicated sources are stored verbatim in reference.

daymade-financial/devils-advocate/references/linqalpha_reference_implementation.md
> "Devil's Advocate is an AI research agent purpose-built to help investors systematically > pressure-test their investment theses using their own trusted sources at 5–10 times the > speed of traditional review."
The reference directory includes the original four-step workflow and production prompt templates (including verbatim JSON schemas) published in the February 2026 AWS blog post. This skill extends them with fields such as base_rate, missing_evidence, and signpost—'how it works' = high-fidelity replication + structured expansion.

The first layer of mechanism explicitly establishes the contrarian persona: since default LLMs tend to flatter the user (RLHF sycophancy), playing devil's advocate must be an explicitly assigned role-based workflow anchored strictly to evidence materials, prohibiting ungrounded nitpicking.

daymade-financial/devils-advocate/SKILL.md
所以反方角色必须被显式赋予，不能指望模型自发唱反调。
SKILL.md uses Sharma et al. ICLR 2024 (sycophancy bias) and Nemeth et al. 2001/2018 (role-play dissent is ineffective; evidence-free contrarianism does more harm than good) as foundational design bases, operationalizing 'authentic dissent' through verbatim citation and evidence anchoring.

Assumption decomposition engine: breaks down the thesis into explicit claims and implicit assumptions (numbered A1/A2), tagged with type (factual/predictive/mechanistic) and load-bearing weight; low-load assumptions are deprioritized and skipped during retrieval; assumption enumeration has no upper limit, while only retrieved items are bounded.

daymade-financial/devils-advocate/SKILL.md
**隐性假设**（implicit assumptions）：没写出来、但论点要成立就必须为真的前提。这一类才是盲点的主产区——确认偏误让人看不见自己依赖了什么。
Implicit assumptions are the prime source of blind spots; the two screening criteria (type dictates what counter-evidence to look for; load-bearing weight dictates whether it is worth retrieving) originate from CIA Key Assumptions Check and RAND ABP.

Counter-evidence retrieval with built-in two-prong questioning and source hierarchy: directly undermining/contradictory materials + ACH-style 'expected-but-absent' evidence; evidence credibility prioritizes official filings > expert calls > broker research > tier-1 media > social media aggregations.

daymade-financial/devils-advocate/SKILL.md
**如果这条假设为真，材料里本应看到、却没看到的证据是什么？**（ACH 的核心一问——缺席的证据也是证据）
The second question is most frequently overlooked: absence of evidence is also evidence. The source credibility hierarchy is adapted from LinqAlpha's original SEC > expert call > broker report order.

Dual-layer artifact delivery: audit layer da_findings.json (structured findings, top-level object to accommodate coverage statements) and human-readable layer da_report.md (thematically grouped narrative + [n] inline citation indices + three lists: References, Surviving Assumptions, Coverage).

daymade-financial/devils-advocate/SKILL.md
默认写入当前工作目录的 `da_findings.json`（用户指定了位置则用用户的；供程序处理与复核）。顶层是一个**对象**而不是裸数组——覆盖面声明要有处安放
The original production prompt used a bare array, whereas this skill switched to an object to include coverage statements; the human-readable layer writes da_report.md by default and delivers it concurrently in the chat interface (Step 4).

Built-in anti-inflation and honesty guardrails: risk_flag evaluates counter-evidence lethality rather than probability; if over half are High, ranking and calibration must be re-evaluated; if no counter-evidence is found, states 'No counter-evidence found in the provided materials' (≠ assumption holds); bias in materials must be explicitly noted in the report.

daymade-financial/devils-advocate/SKILL.md
若超过半数条目被判 High，先把全部条目按杀伤力互相排序再重新定档——全是 High 等于没有分级（首次实测第一版就判出了 6/8 个 High）。
This reflects engineering fixes from the replication benchmark (v1 triggered anti-inflation with 6/8 High flags); 'not found ≠ validated' combined with bias disclaimers (e.g., 'materials are entirely bullish → sparse counter-evidence') jointly uphold evidence integrity.

2Core Capabilities

01Normalize ambiguous theses (prose/IC memo) into a single-sentence claim: 'X will do Y because of Z' and confirm with the user (Step 0)
02Numbered enumeration of explicit/implicit assumptions + load-bearing filtering (unbounded enumeration, load-bearing sets retrieval scope)
03Evidence-anchored counter-arguments: verbatim quotes + source pointers + citations structure, explicitly stating when nothing is found
04Structured findings output (JSON: run_metadata/findings/deferred_assumptions, each assumption containing counter_argument/rebuttal/missing_evidence/base_rate/risk_flag/signpost)
05Base-rate outside view (Mauboussin): populated only when reference classes can be established from the provided materials, strictly prohibiting fabrication
06List of monitorable falsification signals (signpost): extending one-off critique into ongoing monitoring
07Main session fan-out mode: Steps 0-1 run in the main context, then dispatch sub-agents per assumption to retrieve in parallel, save da_findings_A<n>.json to disk, and merge/deduplicate

4Recommended Governance & Evidence

Suggested Controls for Safe Operation
For Agent Runtime
Restrict working directory
For Administrators
Retain operation logs
Evidence & Triggers (File Paths & Excerpts)
  • 本地工作目录: File read/write and artifact output: da_findings.json (current working directory; user can specify alternate path)
Audit Notes: Purely instruction-based skill (directory contains only SKILL.md + 2 reference files, zero scripts/executable code): no script execution, no outbound network requests, no credential access; the only side effects are writing two fixed-name artifacts (da_findings.json / da_report.md, user-configurable) to the current working directory and reading user-provided material files. Considerations: fixed-name overwrites, material scope fully determined by user, narrative phrasing requires human review—none of which pose code-level security risks.

5Second-Pass Verification

  • [ok] 复刻出处（LinqAlpha + AWS 博客）真实存在 — references/linqalpha_reference_implementation.md 收录 AWS ML Blog 2026-02-11 URL、原厂四步流程与生产 prompt 模板；SKILL.md 设计依据表同时列出 CIA/RAND/DEBATE 等来源，无虚构引用。
  • [ok] 『每条反方论证必须带引用 + 逐字摘录』红线 — SKILL.md Step 2 硬规则逐字存在（含省略号/禁改写的细节）；Step 2 置空规则（找不到写『未在给定材料中找到反证』）同在。
  • [ok] 双层产物（da_findings.json 审计层 + da_report.md 人读层） — Step 3 与 Step 4 均声明默认写入当前工作目录；fan-out 子代理落盘 da_findings_A<编号>.json 的路径规则也在 Step 3。
  • [ok] 无脚本/无网络/无凭证（安全结论基础） — 目录仅 SKILL.md + 2 references；全目录扫描无执行/网络/凭证代码；'citations' 字段指向材料文件/URL 而非任何运行时网络调用。
  • [ok] risk_flag 三档锚定 + 防膨胀规则 — 三档定义（High=核心结论不再成立 / Medium=幅度或时间表折损 / Low=边际修正）与『超半数 High 重新排序定档』原文均在。
  • [ok] base_rate 诚实边界（禁止另起检索/编造） — 『只在给定材料内能构建参考类时填写…禁止编一个基础比率，也不为此另起检索』原文存在，防止模型外搜破坏检验边界。
  • [ok] 功能声明 vs 实际能力夸大检查 — description 声称的能力（拆假设、逐条检索带出处反证、风险分级、base_rate、信号标、双层输出、不编造）逐项有对应 Step/红线实现；未声称自动联网检索或判断论点对错。
  • [ok] 元数据（commit/license/stars） — 本地 HEAD == pin d5c4678cb5d4fd6acc9c922690df035dbd33d247；GitHub API：MIT、stars 1385（2026-09-09 实采）、pushed_at 2026-09-09T12:33:29Z。

6Verdict & Scenarios

  • Turns devil's advocacy from an attitude into an auditable process: every counter-argument is anchored to source materials, banning ungrounded arguing.
  • Directly confronts LLM sycophancy: explicitly assigned contrarian role backed by ICLR 2024 / DEBATE 2024 empirical evidence, not merely saying 'please critique'.
  • Blind-spot focus: implicit assumptions and 'expected-but-absent evidence' (ACH) deliver deeper value than surface-level refutations.
  • Disciplined honesty boundaries: explicit when not found, lists unretrieved assumptions, disclaims biased inputs—outputs can be trusted as un-fabricated.
  • Sustainable monitoring: signpost markers transform point-in-time critiques into operational tracking items for ongoing workflows.
    • Best for: Ideal for pre-Investment Committee (IC) / IC memo thesis pressure testing, bear case construction, business decision red-teaming, and research conclusion reviews—especially for analysts and decision-makers who already have evidence materials (research reports, call transcripts, 10-Ks, chat logs) and want 'sourced, auditable, and monitorable' counter-arguments; also fits scenarios requiring proof to teams that counter-arguments were thoroughly examined.
    Not for: Not suitable for scenarios with no provided materials and no willingness to retrieve/build a corpus first (materials-free mode only delivers assumption decomposition + signposts, with reduced coverage); not suitable for users expecting the skill to browse the web autonomously to verify 'thesis vs reality'; not suitable for users seeking overall right/wrong verdicts or buy/sell recommendations (hard boundary: no investment judgments rendered); not suitable for conversations seeking purely emotional debate or argumentative role-play.
    Install Agent-Ready
    ① Verified Mirror 2026-09-09
    Option A · Direct DownloadDownload devils-advocate.tar.gz
    sha256: fafd917ff4bed4b2…
    Option B · Copy Guide for Agent
    install.json
    Agents automatically download and verify mirrors from this JSON guide.
    ② Upstream GitHub
    GitHub Repo ↗
    Pinned to commit d5c4678cb5.