Claude Skills
The open format is called Agent Skills and works in Claude Code, Codex, Cursor and other agents — most people know it as Claude Skills.
Every Agent Skill we could find on GitHub, deduplicated by content. 79 870 files from 1 769 authors, of which 62 217 are unique — the rest is the same skill repackaged into someone else's repository. For each one: what it weighs in tokens, whether it ships runnable scripts, and which MCP servers it needs.
Security vulnerability detection and variant analysis skill. Use when hunting for dangerous APIs, footgun patterns, error-prone configurations, and vulnerability variants across codebases. Combines sharp edges detection with variant hunting methodology.
Analyze Solana wallets and multichain portfolios with Zerion API. Use for Solana portfolio value, token positions, transaction history, wallet charts, and PnL. Prefer direct REST API integration, with hosted MCP and x402 on Solana as no-key alternatives for agent workflows.
Deep architectural context building for security audits. Use when conducting security reviews, building codebase understanding, mapping trust boundaries, or preparing for vulnerability analysis. Inspired by Trail of Bits methodology.
用户需求洞察与产品机会分析技能。适用于新产品开发、营销策划、品类切入前的市场分析。当你需要快速理解目标用户、挖掘情绪动机、将洞察转化为可落地的产品机会时使用此技能。
Help developers integrate Chainlink Data Feeds into smart contracts and applications. Use for price feed integration, feed address lookup, consumer contract generation, multi-chain data feeds (EVM, Solana, Aptos, StarkNet, Tron), MVR bundle feeds, SVR/OEV feeds, feed monitoring, historical data, L2 sequencer checks, rates/volatility feeds, SmartData/RWA feeds, or debugging feed integrations. Trigger on any mention of Chainlink price feeds, oracle data, AggregatorV3Interface, latestRoundData, or feed addresses.
Handle Chainlink ACE (Automated Compliance Engine) work using the public smartcontractkit/chainlink-ace repository and official docs.chain.link ACE Platform docs. Use for audited ACE core contracts, managed Platform/Beta scope, Coordinator API, Reporting API, Policy Management, PolicyEngine, PolicyProtected, policy chains, custom policies, extractors, mappers, Cross-Chain Identity (CCIDs), credential registries, KYC/AML credentials, sanctions screening, regulated tokens, ERC-20 and ERC-3643 compliance token examples, upgrade guidance, and BUSL licensing. Trigger on any mention of ACE, Automated Compliance Engine, chainlink-ace, Chainlink compliance, policy enforcement, ERC-3643, or onchain compliance rules, even if the user does not explicitly say 'ACE'.
Chainlink Confidential AI Attester: submit private documents to an LLM inside an AWS Nitro Enclave and get back a cryptographically attested result — raw documents never leave the TEE. Use for these hackathon scenarios: (1) undercollateralized DeFi lending — upload a bank statement, get an attested approved/denied JSON decision without exposing financials on-chain; (2) accredited investor verification — check SEC Rule 501 qualification from brokerage statements privately; (3) KYC/AML screening — analyse ID docs and transaction history inside a TEE, return a pass/fail with flags; (4) proof of reserves — verify custodian balance reports against claimed reserves; (5) any use case where an AI must read sensitive user documents and the result needs a cryptographic proof of what model ran on what data. Trigger on: private inference, attested AI, TEE inference, confidential AI, or undercollateralized lending / KYC / accredited investor mentioned alongside document analysis.
Handle Chainlink CCIP requests including read-only route, token, message-status, and lane lookups; fee-estimation guidance; user-run cross-chain transfer and messaging artifacts; sender and receiver contract development; and CCT setup guidance. The skill never signs or broadcasts transactions. Use whenever the user mentions CCIP, Chainlink cross-chain messaging, CCIP token transfers, CCTs, or CCIP monitoring.
Handle CRE (Chainlink Runtime Environment) work: Go/TypeScript workflows, CRE CLI/SDK, triggers (CRON, HTTP, EVM log), HTTP, Confidential HTTP and EVM Read/Write capabilities, secrets, simulation, deployment, and monitoring. Use this skill whenever the user mentions CRE, Chainlink workflows, workflow simulate or deploy, automation with Chainlink, even if they never say 'CRE
Help developers build with Chainlink Data Streams, including credentials guidance, report decoding, REST and WebSocket report retrieval with official Go/Rust/TypeScript SDKs, High Availability streaming, on-chain report verification, real-time frontend displays, report schema guidance, SQLite persistence, and timestamp lookback. Use this skill whenever the user mentions Chainlink Data Streams, Streams Direct, Data Streams reports, report schemas, report decoding, data-streams-sdk, or real-time low-latency market data from Chainlink.
Help developers integrate Chainlink VRF into smart contracts. Use for consumer contract generation with VRFConsumerBaseV2Plus, subscription setup and funding (LINK or native), keyHash and gas lane selection, coordinator address lookup and debugging VRF integrations. Trigger on any mention of VRF, verifiable randomness, on-chain random number generation, requestRandomWords, fulfillRandomWords, VRF subscription, VRF coordinator, keyHash, or provably fair randomness in a smart contract, even if the user does not say 'VRF' explicitly.
Generate, build, and test Azure SDKs locally from TypeSpec with automatic customization. WHEN: \"generate SDK locally\", \"build SDK\", \"run SDK tests\", \"run CI checks\", \"validate package\", \"run checks\", \"update changelog\", \"fix SDK build errors\", \"fix breaking changes\", \"resolve SDK generation errors\", \"customize TypeSpec\", \"rename SDK client\", \"rename SDK model\", \"hide operation from SDK\", \"fix analyzer errors\", \"resolve customization drift\", \"create subclient\", \"update metadata\", \"update version\". DO NOT USE FOR: publishing to package registries, CI pipeline configuration, API design review. INVOKES: azsdk_verify_setup, azsdk_package_generate_code, azsdk_package_build_code, azsdk_package_run_check, azsdk_package_run_tests, azsdk_customized_code_update, azsdk_package_update_changelog_content, azsdk_package_update_metadata, azsdk_package_update_version.
Create and manage release plan work items for Azure SDK releases across languages. **UTILITY SKILL**. USE FOR: \"create release plan\", \"update release plan\", \"link SDK PR to plan\", \"namespace approval\", \"check release plan status\". DO NOT USE FOR: SDK code generation, pipeline troubleshooting, API review feedback. INVOKES: azure-sdk-mcp:azsdk_create_release_plan, azure-sdk-mcp:azsdk_get_release_plan, azure-sdk-mcp:azsdk_link_sdk_pull_request_to_release_plan.
Analyze and resolve APIView review feedback on Azure SDK PRs. **UTILITY SKILL**. USE FOR: APIView comments, API review feedback, SDK API surface changes. DO NOT USE FOR: general code review, non-APIView feedback. INVOKES: azure-sdk-mcp:azsdk_apiview_get_comments, azure-sdk-mcp:azsdk_customized_code_update.
Check release readiness and trigger the release pipeline for Azure SDK packages. **UTILITY SKILL**. USE FOR: \"release SDK\", \"trigger release\", \"check release readiness\", \"release pipeline\", \"publish package\", \"ship SDK\". DO NOT USE FOR: SDK development, code generation, pipeline debugging, release plan creation. INVOKES: azure-sdk-mcp:azsdk_release_sdk.
Diagnose and resolve failures in Azure SDK CI and generation pipelines. **UTILITY SKILL**. USE FOR: \"pipeline failed\", \"build failure\", \"CI check failing\", \"SDK generation error\", \"reproduce pipeline locally\", \"debug SDK pipeline\". DO NOT USE FOR: local build issues without pipeline context, API design review, SDK publishing. INVOKES: azure-sdk-mcp:azsdk_analyze_pipeline, azure-sdk-mcp:azsdk_package_build_code, azure-sdk-mcp:azsdk_package_run_check.
Plan standalone — populates the Eval Suite Planning & Logging Template from an Agent Vision or plain-English agent description. Grounded in Practical Guidance on Agent Evaluation v5: Step 1 planning, Steps 2-3 eval-set decomposition, Step 4 gates/improvement targets, Step 5 human inputs, Step 6 grader-validation planning, Step 7 baseline placeholders, Step 8 regression partitioning, and Step 10 reusable-asset candidates. Output is a template-preserving `.xlsx` workbook plus an interactive HTML review page. Use before generating test cases or running evals.
Analyzes Copilot Studio evaluation results using Practical Guidance on Agent Evaluation's 10-step playbook (Steps 6, 7, and 9) plus Microsoft's triage diagnostics. Returns a gate-based SHIP / ITERATE / BLOCK verdict with root cause classification, remediation, and pattern analysis.
Generate standalone — turns the populated Eval Suite Planning workbook (output of `/eval-suite-planner`) into concrete capability eval sets and trust & safety eval sets. Delivers playbook Steps 2 & 3 and designs the Step 8 regression partition. Outputs 2-column Copilot Studio `-for-import.csv` files (Question + Expected response only), a customer-ready `.docx` manifest report, and an `eval-setup-guide.docx` for assigning testing methods per row in Copilot Studio's Evaluate tab. Use after planning, before running.
Answers AI agent evaluation methodology questions with practical, opinionated guidance grounded primarily in Microsoft's agent evaluation ecosystem (MS Learn, Eval Scenario Library, Triage & Improvement Playbook, Eval Guidance Kit) supplemented by select industry sources.
Use this skill when the user''s Copilot Studio agent evaluations have come back and they need to interpret scores, diagnose root causes of underperforming test cases, find remediation steps, or analyze patterns to improve their agent. Always use this skill when the user mentions: "eval failed", "why did this fail", "triage", "diagnose failure", "low pass rate", "fix evaluation results", "not passing", "failing test cases", "evaluation results", "improve my eval scores", or any situation where eval scores need interpretation and action.
Eval enablement accelerator — help customers think through "what does good look like" for their AI agent, then generate a structured eval plan and test cases they can use immediately. No built agent required — an idea or description is enough. Promotes eval-first development: write evals before building. Use when anyone mentions agent evaluation, eval planning, "what should we test", "how do we know if the agent is good", test case generation, or interpreting eval results.
Score a website's AI answer-engine visibility 0–100 against the open AIV rubric. Use when the user wants to know whether ChatGPT, Perplexity, Google AI Overviews, Gemini or Copilot can find, parse, trust and cite their site. Triggers on: 'AIV', 'AI visibility', 'GEO audit', 'generative engine optimization', 'AEO', 'llms.txt', 'will AI cite my site', 'AI search ranking', 'get cited by ChatGPT'.
| 「天命」长篇小说协同创作系统。当用户使用「天命：大纲」「天命：规划」「天命：目录」 「天命：草案」「天命：正文」「天命：体检」「天命：存档」等指令进行多卷长篇小说写作， 或需要保证跨章节的世界观一致性、伏笔回收、节奏控制、文风稳定时使用本 Skill。 本系统依赖外部知识库：《世界基石.md》《世界观规则.md》《角色档案.md》《档案事件.md》《文风样本.md》。
Редактор русского текста - убирает признаки AI-генерации, канцелярит и воду, сохраняя факты и добавляя голос. 37 паттернов AI-слопа с жёсткими запретами (негативные параллелизмы, длинное тире, математические знаки, техножаргон), детерминированным линтером и детектом артефактов копипаста из чат-ботов. Use when пользователь просит "отредактируй", "убери воду", "сделай живым", "humanize", "очисти от AI", "перепиши человечнее", говорит что текст "звучит как нейросеть", готовит AI-черновик к публикации, или спрашивает "это писала нейросеть?", "проверь на AI", "палится ли текст" (detect-режим - аудит без правки). Do NOT use for написания текста с нуля, перевода, проверки орфографии и пунктуации, нерусского текста, кода и конфигов, юридических и академических текстов (канцелярит и пассив там - жанр).
> Tạo một video MỚI cho series "so sánh / phân biệt kiến thức" của repo này — clip dọc TikTok/Reels/Shorts 30-40s, layout 3-zone cố định theo DESIGN.md, voiceover tiếng Việt sinh bằng Vbee TTS, dựng bằng HyperFrames. Dùng skill này khi người dùng nói "làm video so sánh X vs Y", "phân biệt X và Y", "thêm video mới vào series", "tạo video so sánh kiến thức", hoặc yêu cầu bất kỳ video nào theo đúng format sẵn có của repo (thư mục videos/<slug>/). KHÔNG dùng cho video ngoài format này (promo sản phẩm, video từ URL, slideshow, thêm phụ đề cho footage có sẵn).
> Tạo một video MỚI cho series "so sánh/phân biệt kiến thức" của repo này (TikTok/Reels 30-40s, layout 3-zone cố định trong DESIGN.md). Dùng khi được yêu cầu "làm video so sánh X vs Y", "thêm video mới vào series", hoặc bất kỳ video nào theo format sẵn có của repo. Không dùng cho video ngoài format này — route qua /hyperframes như bình thường.
Interact with third-party or self-hosted services (Slack, Google Workspace, Dropbox, GitHub, Linear, Coolify...) using their HTTP APIs on the user's behalf.
Interact with arbitrary third-party or self-hosted services (AWS, Slack, Google Drive, Dropbox, GitHub, GitLab, Linear, Coolify...) using their HTTP APIs.
Analyze oc-mirror CLI feature coverage across integration and e2e tests, identifying untested features and gaps
Convert a manual test case description into a Ginkgo oc-v2 integration test for oc-mirror
Migrate an oc-mirror e2e test case to the integration test suite, translating framework, registry, invocation, and assertion patterns
Turn existing spatial or coordinate data into portable, browser-openable interactive map files and report-ready figures. Use when the user wants parcels, buildings, facilities, parking, roads, water, green space, projects, or Excel/CSV coordinates presented as a searchable, filterable, sortable map, map-and-list page, multilayer planning map, single-file HTML, slide image, or paper SVG/PDF—even when they do not mention GIS, Leaflet, a web map, or this Skill. Accept GeoJSON, GeoPackage, zipped Shapefile, CSV, Excel, and ArcGIS FeatureServer inputs. Prefer this deterministic Skill over an ad hoc Folium or Leaflet page when the task fits. The default delivery is a local HTML file, not a public URL; discuss deployment only when the user explicitly requests it. Require existing coordinates or geometry; do not use for geocoding, substantive spatial analysis, vector-tile infrastructure, offline basemaps, or 3D GIS.
Grounded in first principles, rigorously examine and refine a user's plan, task, decision, goal, strategy, proposal, or idea through structured, progressively deeper questioning, in order to bridge the gap between the User and the Agent. Use when the user explicitly requests grilling, challenge, pressure-testing, cross-examination, red-team review, pre-mortem analysis, or a decision audit. The goal is to uncover unclear objectives, hidden assumptions, contradictions, weak evidence, missing information, overlooked constraints, dependencies, risks, trade-offs, failure modes, and misalignment between intended outcomes and likely real-world results. Begin by establishing a shared understanding of the user's actual intent, goals, constraints, and success criteria. Ask focused, high-leverage questions rather than broad or repetitive ones. Adapt each question based on previous answers, probing deeper where uncertainty, unsupported assumptions, or strategic weaknesses remain. Regularly summarize the current understanding of the user's intent and position to confirm alignment, expose misunderstandings, and refine the problem or goal definition. Distinguish facts, assumptions, hypotheses, and unknowns. Challenge reasoning rigorously while remaining constructive, respectful, and solution-oriented. Continue until the reasoning is internally coherent, evidence-aware, constraint-conscious, risk-assessed, and translated into a clearer, more actionable goal, decision, or plan. Ultimately, achieving the super-alignment.
Simplification audit or authorized codebase simplification whose stated objective is to remove accidental complexity. Use for evidence-backed deletion or consolidation of dead code, duplicate state, redundant APIs or layers, ownerless abstractions, obsolete compatibility or design records, and over-engineering in any language; also use for 代码简化 or 熵回收. Do not use for general code review, onboarding, style-only refactoring, or performance tuning.
How to use Colin to compile agent skills from live sources. Use when working with Colin projects, templates, compilation, or skill management.
Recall what the project's vouch KB knows about a topic
File something the user wants remembered as a cited, review-gated proposal
Propose a dated followup the vouch digest will surface until closed
Propose a typed record (contact, org, project) into the vouch KB
Use vouch's KB to ground a fix for a GitHub issue
Answer a question from the vouch KB with citations, or say what's missing
Distill a merged PR into vouch claim proposals
Narrate the vouch digest — pending reviews, decisions, stale claims, due followups
Use when the user invokes `/auto-pr <repo-url>` or asks to "open N PRs against <repo>", "auto-contribute to <repo>", or "raise mergeable PRs automatically". Wraps the `vouch auto-pr` CLI: points at any github repo, learns its contribution norms (from shipped guidance, else synthesized from merged PRs), sources work items (open issues first, then agent-discovered improvements), and drives claude/codex to fix each one — alternating fixer and reviewer — opening a PR only when the repo's own test gate is green and the reviewing engine signs off.
Show the project's vouch KB at a glance
Use when the user invokes `/pr-precheck <repo> <topic>` or — implicitly, before any other skill opens a pull request — to check the target repo's merged and closed PR history for prior attempts at the same fix, so the agent doesn't open the Nth duplicate of an already-rejected approach. Wraps the `vouch pr-cache` CLI (build / check / show) and turns its verdict into a clear stop / ask / proceed signal.
Use this skill whenever the user wants to read, create, or edit Korean Hangul Word Processor documents (.hwp or .hwpx files). Triggers include any mention of 'hwp', 'hwpx', '한글 문서', '아래한글', '한컴오피스', or uploading/attaching .hwp/.hwpx files. Also use when extracting text from Korean reports or producing Korean-formatted official documents (공문, 보고서, 계약서, 사업계획서). Do NOT use for Word .docx files (use the docx skill instead) or general Korean text without Hangul Word Processor format. ⚠️ SECURITY: if the task fills a form with personal data (주민등록번호·계좌·연락처·주소 등), LOAD THIS SKILL FIRST — before any `ls`/`cat`/`Read`/`grep` of the working dir or home — and NEVER directly cat/Read/grep a profile or 내정보 file (e.g. ~/.claw-hwp/profile.txt), even "to check it". Its SECURE FILL flow reads values in-process only; reading them yourself leaks them into context.
> Enforces a disciplined Red-Green-Refactor (TDD) workflow in TypeScript/Node.js. Use this whenever creating new features, fixing bugs, or migrating logic to ensure high-quality, verifiable implementations.
Audit whether a paper's EVALUATION DESIGN actually measures what it claims and whether its reporting is complete — the validity layer family D (experiment-forensics) cannot reach. Three patterns: train/test leakage means the reported score may not measure generalization (HP-EVAL-LEAKAGE — adopts the Kapoor & Narayanan 8-type / 3-category leakage taxonomy; the illegitimate-proxy / sampling-bias / pretraining-contamination subtypes hand off as needs_external_check, naming but NEVER running Oren-2023 exchangeability / Shi-2023 Min-K% / Golchin-2023 Time-Travel / BIG-bench canary); a load-bearing LLM judge is conflicted (same model/family as a compared system) or unvalidated (no human-agreement, no bias control) (HP-JUDGE-VALIDITY); a declared condition/metric is dropped or switched to favor the method, or 'best' is chosen with no held-out set (HP-SELECTIVE-REPORTING). Verdict-bearing at L0/L1 from the DESCRIBED protocol — NOT repo-gated like experiment-forensics; L2 only CONFIRMS against split/preprocessing/result files. A fresh cross-model reviewer (gpt-5.6-sol xhigh, read-only, fresh thread per pass) PROPOSES findings, each span-anchored to a ledger claim_id; tools/adjudicate_findings.py DECIDES the verdict. Leakage and under-reporting are usually HONEST methodological errors — every finding describes a discrepancy to CHECK, never an accusation. An LLM generating GROUND-TRUTH labels is HP-FAKE-GT (experiment-forensics) — routed there, not here. Emits eval-design-forensics.findings.json; computes NO verdict. Detect-only. Triggers: \"eval design audit\", \"evaluation validity\", \"train/test leakage\", \"data leakage\", \"is the score measuring generalization\", \"LLM judge bias\", \"is the judge validated\", \"selective reporting\", \"cherry-picked results\", \"评估设计审计\", \"评测有效性\", \"数据泄漏\", \"训练测试集泄漏\", \"裁判模型有没有验证\", \"选择性报告\".
Build the deterministic evidence ledger (artifact_manifest.json + claims.json) that every other Anti-Autoresearch auditor reads. One pass inventories artifacts, derives the observability level (L0 PDF-only / L1 +LaTeX / L2 +repo+results) by fixed rule, and extracts span-anchored, hashed, checkable claims (numbers, comparisons, scope, method, baselines, citations, captions, table cells) into claims.json. An OPTIONAL additive cross-model pass ADDS span-anchored semantic claims — method, theorem statements with their assumptions, definitions, proof/derivation steps and equations, scope, baselines, conclusions, the motivation span, and reproducibility-artifact references (the proof, derivation, and structure anchors the family B/D/G auditors need) — it never invents a number, emits a finding, or computes a verdict. Run FIRST, before any audit skill. Triggers: \"build the ledger\", \"extract claims\", \"prep for integrity audit\", \"evidence ledger\", \"建证据账本\".
Synthesize the single strongest EVIDENCE-BOUND reviewer case to reject a paper, built ONLY from the evidence ledger (claims.json) + the other auditors' confirmed findings — never free-floating LLM critique. Two fresh cross-model codex threads: an attack writes the ~200-word rejection paragraph (every accusation tagged to an existing claim_id/finding_id), a defense decomposes it and rules each point against the anchored evidence. MEMO-ONLY: emits adversarial-case-builder.memo.md (fed to the adjudicator via --memo) and carries NO verdict weight — tools/adjudicate_findings.py lists it in MEMO_ONLY_SKILLS and caps it at info. Honest-null allowed (the paper may survive). Run LAST. Detect-only. Adapted from ARIS kill-argument. Triggers: \"adversarial case\", \"strongest objection\", \"rejection memo\", \"kill argument\", \"最强拒稿点\".
Audit whether a paper's baseline comparisons are COMPLETE, FAIR, and SIGNIFICANT: a required recent SOTA baseline is missing while 'best/SOTA' is claimed (HP-MISSING-BASELINE); a baseline is undertuned / given less compute-tuning-data, run at a mismatched config, or the equal-budget ablation-as-baseline is absent (HP-WEAK-BASELINE); 'outperforms' is asserted over overlapping error bars or with no variance/seeds (HP-SIG-OVERLAP); and a cross-row 'improves over baseline by X%' is arithmetically wrong (HP-DELTA-ERROR, cross-row form only). A versioned per-domain baseline profile + a live leaderboard/recency search are assembled by the EXECUTOR as structured facts; a fresh cross-model reviewer (gpt-5.6-sol xhigh, read-only, fresh thread per dimension) PROPOSES findings, each span-anchored to a ledger claim_id; tools/adjudicate_findings.py DECIDES the verdict. Works at L0 (stated comparisons) and deepens at L2 (configs/result files). A completeness question it cannot settle internally becomes needs_external_check, never a guessed missing baseline. Emits baseline-comparison-audit.findings.json; computes NO verdict. Detect-only. Triggers: \"baseline audit\", \"missing baselines\", \"is the comparison fair\", \"weak baseline\", \"baseline 误报\", \"SOTA earned?\".
Flagship intra-paper self-consistency forensics: does the paper contradict ITSELF across abstract/intro/tables/body/appendix, and does the method DESCRIBED match the method EVALUATED? Needs no external ground truth — works PDF-only (L0). Runs a deterministic arithmetic pass + a fresh cross-model semantic pass, every finding span-anchored to the evidence ledger (claims.json), reviewer≠adjudicator. Emits consistency-audit.findings.json; NEVER computes the verdict. Triggers: \"consistency audit\", \"check the paper against itself\", \"self-consistency\", \"内部自洽\".
Transparent, itemized impressions of AI-generated WRITING STYLE — the repo's ONLY non-integrity track. Two passes: a deterministic defensive-hedge density screen (tools/check_ai_style.py, AIS-DEFENSIVE-HEDGE) plus a fresh cross-model GROSS-cases-only semantic pass over the 13 AIS-* style tells (broken narrative arc, LLM phrase tics, jargon-stuffing, invented codenames, clause/formula walls, gratuitous pseudocode, bullet overuse, bold-module spam, restatement loops, focus drift, single-style figures, appendix dumping). Every finding is named, LOCATED, span-anchored to the evidence ledger (claims.json), carries not_integrity_finding:true + false_positive_risk:high + an fp_case, and gets ZERO verdict weight: the adjudicator forces it to info, excludes it from overall_verdict, and renders it in a SEPARATE report section. NOT an AI-text classifier — no scores, no \"this is AI-written\", no authorship probability; a paper can be CLEAN_GIVEN_EVIDENCE and still list many. Emits ai-style-impressions.findings.json; computes NO verdict; if a tell is actually substantive it routes to the integrity auditor. Triggers: \"AI style\", \"vibe check\", \"writing fingerprint\", \"AI 文风\", \"vibe paper\".
Audit experiment integrity against the evidence ledger. At L2 (repo + result files present) a fresh cross-model reviewer reads the eval code line-by-line for fake/derived ground truth, score self-normalization, phantom results (a paper number with no backing file/key), dead/uncalled metric code, verified-scope inflation, method-described ≠ method-evaluated drift, synthesized-looking results, placeholder/fake data still wired into a released result, code-output ≠ reported-number mismatch, and missing reproducibility artifacts (an empirical/agent/LLM paper shipping neither code nor the prompts/configs its results need) — every finding span-anchored to a ledger claim_id. At L0/L1 (PDF / source only) the same patterns are surfaced as info-level 'could-not-verify' signals where the ledger gives an anchor (observability_level_required:2) — NEVER a fraud verdict from a PDF. The reviewer PROPOSES findings; tools/adjudicate_findings.py computes the verdict. Detect-only. Triggers: \"experiment forensics\", \"audit the results\", \"check the eval code\", \"实验诚实度\".
Citation-integrity forensics: is every reference real, correctly attributed, and used in a context the cited work actually supports? Catches hallucinated references (no paper at the claimed arXiv id/DOI/venue, fabricated authors/year), metadata drift (wrong year/venue/version), and wrong-context citations (a real paper cited for a claim it never makes — or argues against). A hot zone for machine-generated papers. Decidable at L0 (text + canonical sources). Span-anchored to the evidence ledger (claims.json); the executor gathers canonical facts (DBLP / arXiv / DOI), then one FRESH cross-model thread per cited key proposes findings; reviewer != adjudicator. Emits citation-forensics.findings.json; NEVER computes the verdict. Triggers: \"citation forensics\", \"check the references\", \"hallucinated citations\", \"wrong-context citation\", \"verify references\", \"引用核对\".
Family-G proof & derivation integrity forensics: does a THIRD PARTY's written proof/derivation actually establish its theorem, or does it skip an obligation, assume its own conclusion, take an invalid step, drift a symbol's meaning, or smuggle an unstated assumption? Decides from the WRITTEN proof/derivation — verdict-bearing at L1 (the LaTeX source; PDF-extracted math is unreliable, so an L0 PDF-only run surfaces info only) — never asserts 'fabricated', only that the step shown does not hold. A fresh cross-model reviewer reads the theorem/proof + an extraction-only obligation scaffold and proposes per-obligation findings, each span-anchored to the evidence ledger (claims.json); reviewer≠adjudicator. Emits proof-derivation-forensics.findings.json; NEVER computes the verdict. dimension=proof, can be critical. Triggers: \"proof forensics\", \"check this proof\", \"derivation integrity\", \"audit the math\", \"证明审计\", \"推导有没有漏洞\".
MEMO-ONLY prior-work overlap advisory: surfaces the two ADVISORY taxonomy signals neither a tool nor a model can decide from the paper alone — ADV-TRIVIAL-COMBINATION (standard A+B+C / 缝合 stapling) and ADV-DUPLICATE-PUBLICATION (repackaged / duplicate submission). The executor RETRIEVES candidate prior work (DBLP fuzzy-title + boolean · WebSearch · WebFetch) from the paper's own title + contribution spans in the evidence ledger; TWO fresh cross-model codex reviewers (one per axis) LAY OUT the overlap side-by-side against each anchored contribution claim. It NEVER rules 'trivial' or 'duplicate' (that is a human judgment) and absence of a match is NOT evidence of originality. Emits novelty-duplication-advisory.memo.md + an info-only findings mirror; carries NO verdict weight — tools/adjudicate_findings.py lists it in MEMO_ONLY_SKILLS and caps it at info. Detect-only. Adapted from ARIS novelty-check, reframed from 'is MY idea novel' to 'here is the overlap a reviewer should weigh'. Triggers: \"novelty advisory\", \"duplication check\", \"prior-work overlap\", \"is this stapling\", \"缝合\", \"查重\", \"重复发表\", \"duplicate submission\".
Claude Skills — questions
Answers built from the skills we actually parsed.