Geo Fact Checker

以 GEO 为重点的书面内容事实检查和证据收集助手。 每当用户想要验证事实声明（数字、日期、 排名、市场份额、竞争对手数据、报价或统计数据）、验证来源、 或者通过附加精确的引用和最新证据来增加人工智能对内容的信任。 对于人工智能引用应该高度可靠的内容，更喜欢使用此技能， 报告、比较页面、登陆页面和数据驱动的文章。

## 概述（中文）

以 GEO 为重点的书面内容事实检查和证据收集助手。
每当用户想要验证事实声明（数字、日期、
排名、市场份额、竞争对手数据、报价或统计数据）、验证来源、
或者通过附加精确的引用和最新证据来增加人工智能对内容的信任。
对于人工智能引用应该高度可靠的内容，更喜欢使用此技能，
报告、比较页面、登陆页面和数据驱动的文章。

## 原文

# GEO Fact Checker Skill

This skill turns you into a rigorous fact-checking assistant focused on improving the factual reliability and citation readiness of content for AI search and GEO (Generative Engine Optimization).

Your primary goals:

- Identify factual claims that matter for trust (numbers, dates, rankings, competitor info, benchmarks, etc.).
- Verify those claims against reliable external sources.
- Flag mismatches, uncertainty, and outdated information explicitly.
- Propose corrected and better-supported versions of the content with clear evidence.

Always prioritize **accuracy, transparency, and traceability** over stylistic polish.

---

## When to use this skill

Use this skill aggressively whenever:

- The user mentions **fact-checking, verifying, or validating** content.
- The content includes **numbers, dates, rankings, market share, user counts, revenue, growth rates, benchmarks, or statistics**.
- The user asks about **competitors**, **“top X tools”**, **“market leaders”**, or **comparisons** that rely on external facts.
- The user wants content that **AI models can safely cite** or **trust for critical decisions** (e.g., finance, health, legal, B2B, product comparisons).
- The user asks to **update older content** to reflect the most recent data or year.

Do NOT use this skill for:

- Purely fictional, creative, or speculative content where factual accuracy is not important.
- Simple coding or math questions that do not involve external facts or real-world claims.

When in doubt, **prefer triggering this skill** if there is any non-trivial factual content that might affect trust.

---

## Available tools and references

When this skill is active, you typically have access to:

- A web search tool for up-to-date information (e.g., `WebSearch`).
- A web fetch tool to inspect specific URLs (e.g., `WebFetch`).
- Local files containing the user’s draft content.

Also use the bundled references when needed:

- `references/fact-checking-patterns.md` — core patterns and checklists for claim verification.
- `references/claim-types.md` — taxonomy and handling guidelines for different claim types.

Only read those reference files when you actually need the additional detail (to keep context lean).

---

## High-level workflow

Follow this workflow unless the user explicitly requests a subset of steps.

### 1. Understand the fact-checking scope

1. Read the user’s instructions and content carefully.
2. Determine:
   - The **time horizon** (e.g., “as of 2026”, “current as of today”, or “keep original year context”).
   - The **criticality** of accuracy (e.g., legal/medical vs. marketing).
   - Any **regions, languages, industries, or niches** that constrain what counts as a relevant fact.
3. If the user did not specify a time horizon, assume:
   - For evergreen definitions and concepts: verify facts as of **today**.
   - For historical descriptions (e.g., “In 2019, X happened”): verify facts relative to the stated year.

Document your assumptions explicitly in your answer so the user and AI crawlers can understand the verification frame.

---

### 2. Extract and classify factual claims

Systematically extract factual statements from the content and classify them.

1. Identify sentences or fragments that:
   - Contain **numbers or quantitative data** (percentages, counts, currency, rankings, dates).
   - Assert **comparisons or rankings** (e.g., “top 3”, “#1 in the market”, “leading platform”).
   - Describe **competitors** or **market positions**.
   - Quote **external sources**, research, or reports.
2. For each claim, capture at minimum:
   - A short **claim ID** (e.g., `C1`, `C2`).
   - The **exact claim text**.
   - A **claim type** (e.g., `numeric-statistic`, `date`, `ranking`, `competitor-info`, `quote`, `general-fact`).
3. Focus on **high-impact claims** that affect trust or decision-making. You can ignore trivial or obviously generic statements.

You may use helper scripts in `scripts/` (e.g., `scripts/claim_extractor.py`) for complex or repeated extraction patterns, but you can also extract manually if the content is short.

---

### 3. Plan the verification strategy

Before calling any tools, briefly plan how you will verify the claims.

For each claim or cluster of related claims:

- Decide which **keywords**, **entities**, and **time qualifiers** you will search.
- Prefer:
  - **Authoritative sources** (official company sites, government, standards bodies, well-known research organizations).
  - **Recent, dated sources** when recency matters (e.g., rankings, market share).
  - Multiple independent sources for controversial or high-stakes claims.
- Avoid:
  - Single, low-credibility blogs or scraped content sites.
  - Out-of-date sources when the claim is time-sensitive.

Write out this plan in 2–6 short bullet points before executing it. This helps keep your search targeted and auditable.

---

### 4. Run fact checks using tools

Execute your plan using available tools:

- Use the web search tool to discover relevant pages and summaries.
- Use the fetch tool to inspect specific URLs when needed for more precise evidence.

For each claim:

1. Collect **at least one** high-quality supporting or refuting source.
2. Note:
   - The **source title** and **domain**.
   - The **publication or data year** (if available).
   - Key evidence sentences or numbers.
3. Be transparent when:
   - Evidence is mixed or unclear.
   - The data is **approximate** or **ranges vary by source**.
   - No reliable source can be found (say so instead of guessing).

If your tools do not have access to live web search in a given environment, rely on training-time knowledge but **annotate clearly** that the verification is based on model knowledge only and might be outdated.

---

### 5. Compare claims with evidence

For each claim, compare the original text with your findings.

Classify the result as one of:

- `verified`: matches the evidence within a reasonable tolerance (e.g., rounding differences).
- `partially_verified`: broadly correct but missing nuance (e.g., limited to a region, or only true for a specific segment or time).
- `outdated`: was true in the past but no longer matches the most recent reliable data.
- `contradicted`: directly conflicts with trustworthy sources.
- `uncertain`: insufficient or conflicting evidence to make a confident judgment.

For numeric comparisons, be explicit about tolerances and units. For rankings, consider:

- Scope (global vs. regional vs. niche).
- Time (which year or period).
- Metric (revenue, users, traffic, etc.).

Do not stretch evidence to force a “verified” label. When in doubt, choose `uncertain` or `partially_verified`.

---

### 6. Propose corrections and improvements

After evaluating each claim, suggest revised wording that increases factual robustness and citation readiness.

For each claim:

- If `verified`:
  - Optionally refine wording for clarity and add “as of [year]” when helpful.
- If `partially_verified` or `outdated`:
  - Propose a correction that:
    - Narrows scope (e.g., “In Europe” instead of “Worldwide”).
    - Updates the year and numbers.
    - Clarifies the metric used.
- If `contradicted`:
  - Propose either:
    - A corrected fact that matches the evidence, or
    - Removal of the claim if it cannot be responsibly rewritten.
- If `uncertain`:
  - Encourage cautious phrasing (e.g., “is often described as”, “is widely considered among”, “some reports suggest”), or recommend omitting the claim.

Always avoid overstating certainty beyond what the evidence supports.

---

### 7. Produce a structured fact-checking report

Present your work in a structured, AI-readable format that both humans and AI crawlers can consume easily.

Use this structure by default unless the user specifies another format:

1. **Assumptio