Docs & Knowledge · daymade/claude-code-skills
fact-checker
Verifies factual claims in documents using web search and official sources, then proposes corrections with user confirmation. Use when the user asks to fact-check, verify information, validate claims, check accuracy, or update outdated information in documents. Supports AI model specs, technical documentation, statistics, and general factual statements.
L3 Networked · Declare domain & key scopeV1 Commit-PinnedNetwork AccessWrite FileDual-pass Audit
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
Pure prompt orchestration with zero scripts: The entire skill consists only of SKILL.md (workflow and specifications) + README.md (external documentation). Fulfillment relies entirely on the model executing a 5-step checklist.
fact-checker/SKILL.md
Verify factual claims in documents and propose corrections backed by authoritative sources.
A directory glob confirms only these two files exist—no scripts/, no references/, and no evals/. Therefore, the underlying implementation is purely prompt engineering: encoding the verification workflow into a reproducible checklist (Steps 1–5) and formulating disciplines (when to stop, when to ask, how to cite sources) as hard constraints.
Step 1 begins with 'verifiability screening': it only accepts technical specifications, version numbers/release dates, statistical figures, API capabilities, and benchmark scores, strictly excluding subjective content.
fact-checker/SKILL.md
**Skip subjective content:**
It explicitly lists exclusions: Opinions and recommendations / Explanatory prose / Tutorial instructions / Architectural discussions. This step ensures the skill does not get bogged down in unresolvable debates over whether opinions can be verified, establishing the prerequisite for delivering within scope.
Comparison results are represented by a 4-tier status code, treating 'unable to find an authoritative source' as a first-class citizen status (❓ Unverifiable) rather than forcing the model into a binary choice.
fact-checker/SKILL.md
| Claude 3.5 Sonnet: 200K tokens | Claude Sonnet 4.5: 200K tokens | ❌ Outdated model name | platform.claude.com/docs |
The four tiers defined in the same section are: ✅ Accurate / ❌ Incorrect / ⚠️ Outdated / ❓ Unverifiable. Explicitly modeling 'unable to verify' is a crucial design choice to avoid hallucinations—when no source is found, it is required to output Unverifiable along with qualified phrasing rather than fabricating a source.
The correction report follows a fixed template (Location / Current claim / Correction / Source / Rationale), where every proposed modification must reference a line number and source URL.
fact-checker/SKILL.md
**Location:** Line 77-80 in docs/file.md
In the same template, Correction is immediately followed by **Source:** and **Rationale:** (SKILL.md provides a full example later, including source links from platform.claude.com and openai.com). This discipline makes the justification for a change a mandatory field in the report.
A strict manual gate is enforced before modifications: the report is presented first, explicitly asking 'Should I apply these corrections?', and actions are taken only after receiving an affirmative response.
fact-checker/SKILL.md
2. Wait for explicit approval: "Should I apply these corrections?"
What is happening here: converting write operations from 'model discretion' to 'explicit user approval'. Condition 3 in the Step 5 prerequisite list is 'Only proceed after confirmation'—the skill will never modify user documents without approval.
Temporal context is treated as a mandatory requirement: corrections must include an 'as-of' date, and terms that become instantly obsolete like 'latest version' are forbidden.
fact-checker/SKILL.md
- "Latest version" (becomes outdated)
The contrast section provides positive examples like '"As of January 2026"' and '"Claude Sonnet 4.5 (released September 2025)"'. The Real-World Example in README also lists 'Added temporal marker "As of January 2026"' as one of the edits—aligning the documentation with SKILL.
Authoritative hierarchy of sources is clearly ranked, providing actionable steps to 're-check against official sources' for potentially unreliable sources.
fact-checker/SKILL.md
- Third-party aggregators (llm-stats.com, etc.) - verify against official sources
The authoritative hierarchy is: Official product pages > API documentation > Official blog announcements > GitHub releases (open source). Cross-checking is required for third-party aggregators and blog posts; outdated docs, unsourced wikis, and rumors are directly categorized as Avoid.
Search queries themselves follow explicit guidelines: requiring specificity and inclusion of the year, with generic queries like 'Claude context' cited as counter-examples.
fact-checker/SKILL.md
- "Claude Opus 4.5 context window 2026"
The same section contrasts good and bad queries (Good: Model name + spec + year; Bad: AI models / Latest version) and outlines three rules under Search strategy (use model name + spec, include the current year, and verify across multiple sources whenever possible).
Exit workflows exist for both conflicts and missing information: resolving conflicts by 'prioritizing the latest official doc + noting discrepancies in the report + presenting both to the user', and providing qualifying phrasing rather than conclusions when no source is found.
fact-checker/SKILL.md
2. Suggest alternative phrasing: "According to [Source] as of [Date]..."
In this section (Handling ambiguity), rule 3 under 'When no source found' advises adding qualifications (approximately / reported as). This translates uncertainty into safe, guarded wording rather than asserting definitive claims.
Concludes with downstream handoff: after verification, it suggests triggering /daymade-docs export skills to establish a 'verify → publish' pipeline.
fact-checker/SKILL.md
A) Export as PDF — run /daymade-docs:pdf-creator (Recommended for formal documents)
This represents soft orchestration across the daymade-docs plugin family in the same repository (offered merely as suggested commands for the user, not a hard inter-skill dependency).
2Core Capabilities
01Identify verifiable factual assertions in documents (technical specifications, versions, statistics, API capabilities, benchmark scores)
02Search official sources ranked by authoritativeness (Official product pages > API documentation > Official blogs > GitHub releases)
03Generate structured correction reports with line numbers, current claims, proposed corrections, sources, and rationales
04Classify findings into 4 status codes (Accurate, Incorrect, Outdated, Unverifiable), preserving 'Unverifiable' as a valid conclusion
05Apply document edits using the Edit tool only after explicit user approval, followed by commit reminders
06Enforce as-of dates and qualifying phrasing on time-sensitive content to prevent rapid information decay
07Align numerical precision with the source (write 'approximately' if the source says approximately; provide exact values only if the source does)
08Suggest downstream exports (PDF / slides) upon verification completion to complete the publishing pipeline
3External Dependencies
| Kind | Dependency / Endpoint |
|---|---|
| api | 宿主 web search 能力（description 明写 using web search；skill 自身无代码、无端点常量） |
| api | 宿主 Edit 工具（Step 5 实际改写用户文档） |
| network | AI 厂商官方文档/公告域名（作为检索目标列出：anthropic.com、openai.com、blog.google、platform.claude.com、platform.openai.com） |
| network | 包注册表（校验库版本是否过时：npm、PyPI、crates.io；另指向 github.com 的 releases） |
| network | 第三方聚合站（被明确要求「严格管控、回官方核对」） |
| api | 同仓下游插件 daymade-docs（仅作为建议用户运行的命令，非本 skill 调用） |
4Recommended Governance & Evidence
Suggested Controls for Safe Operation
For Agent Runtime
Declare allowed domains and key scope
For Administrators
Least privilege keys, retain request logs
Evidence & Triggers (File Paths & Excerpts)
- 网络端点: Declared external network access: All network traffic is initiated via the host's web search tool (the skill contains no code or endpoint constants of its own); SKILL.md designates search targets as official vendor domains and package registries.
- 本地工作目录: File read/write and artifact output: Modifies assertions in user documents deemed incorrect/outdated using the host's Edit tool (Step 5; strictly conditional on obtaining explicit user approval).
Audit Notes: Based on standard tiering: The entire directory contains only two text files (SKILL.md + README.md), involving no script execution, credential access (no access to env vars, keychains, secrets, or configuration files), and no autonomous writes. Its sole substantive behavior is directing the agent to perform searches and present a correction checklist for user review; document modifications occur only after explicit user consent, protected by built-in manual gates (the three prerequisites in Step 5) and the fallback mechanism of labeling missing sources as 'Unverifiable'. Search targets are constrained to official vendor domains and public package registries, making outbound traffic predictable. Note: Operational risk stems not from system attack surface but from output accuracy—refer to verdict.risks.
5Second-Pass Verification
- [ok] 是否真的没有脚本（description 提到 web search 会不会误导） — 事实核查：目录内无 .py/.sh/.mjs/.js，无 package.json；description 的『using web search』指的是宿主提供的搜索能力，不是自带脚本。故 scripts_executed = []，与 README『No configuration required. The skill works out of the box.』一致。
- [ok] 『用户批准才改』是否只是口号 — SKILL.md Step 5 明确列三条前置（Show the correction report / Wait for explicit approval / Only proceed after confirmation），并给出询问原句 'Should I apply these corrections?'。这是提示词级约束，无技术强制，但表述是命令式而非建议式。
- [discrepancy] 外部资源条目：域名是否被误当成调用点 — 〔第二遍标注：需精确措辞〕SKILL.md 里的 anthropic.com / openai.com / blog.google / platform.claude.com / npm / PyPI / crates.io / llm-stats.com 全部出现在「检索目标示例」与「来源评估」段落中，是给模型的指导清单，不是代码常量。已在 external_deps 的 name 字段逐条写明性质（『作为检索目标列出』『被明确要求严格管控』），避免读者误以为 skill 会主动请求这些站点。
- [ok] 『无法核实』是否真的一等公民（找反例） — 四档里含 ❓ Unverifiable，Handling ambiguity 段落给出无来源时的三条动作（标 Unverifiable / 给限定话术 / 建议加 approximately 之类的限定词），Limitations 段落重申不能判定争议事实。三处互相印证，非单点承诺。
- [ok] 时效性纪律是否两文档一致 — SKILL.md 的 Time-sensitive information 段落给出 Good/Poor corrections 对照；README 的 Real-World Example 把『Added temporal marker "截至 2026 年 1 月"』写进改动清单。两处一致。
- [ok] 数字精度要求是否有对应实现路径 — Numerical precision 段落给出 Source says → Write 的对照（approximately 1 million tokens → 1M tokens (approximately)），并在 Citation format 段落要求把来源写进修正文本。属文档级规范，无脚本可验证，属『提示词约束』性质。
- [ok] 元数据（license/stars/last_push） — 取自同 repo 同 commit 的池内既有值（MIT、1385★、2026-09-09T12:33:29Z）；本次未联网复核 gh。README 自述 License 为 MIT License - See repository for details，与仓库根 LICENSE 一致。
- [ok] 下游交接是否造成隐式依赖 — Next Step 段落只提出选项并要求用户运行 /daymade-docs:pdf-creator 或 :ppt-creator，skill 自身不调用、不 import，属软建议而非硬依赖。
6Verdict & Scenarios
Best for: Ideal for authors of long-form documents, manuscripts, and technical documentation to run periodic specification audits (especially rapidly changing fields like AI model parameters, library versions, and statistics); also suitable for pre-publication checks of team documentation to surface 'outdated but sourced' discrepancies and make item-by-item decisions on revisions. Its structured output naturally serves as review material for change sets, as reviewers can verify each edit directly via line numbers and source URLs.
Not for: Not suitable for scenarios requiring deterministic correctness (e.g., regulatory compliance, legal, or medical conclusions)—it offers 'sourced correction suggestions', not guarantees; not suitable for verifying opinions, subjective judgments, or design trade-offs (which SKILL.md explicitly categorizes under 'Skip subjective content'); not suitable for environments lacking web search capabilities (the entire workflow relies on web search); not suitable for scenarios requiring unattended, fully automated editing without human intervention (every modification requires user approval); and not suitable for verifications requiring access behind paywalls or internal intranets.
Install Agent-Ready
① Verified Mirror 2026-09-15
Option A · Direct DownloadDownload fact-checker.tar.gz
sha256:
Option B · Copy Guide for Agentfedd09ed5369f915…Agents automatically download and verify mirrors from this JSON guide.
② Upstream GitHub
GitHub Repo ↗
Pinned to commit d5c4678cb5.