| name | extract-facts |
| version | 1.0.0 |
| description | Extract discrete, verifiable facts from web content with source citations. |
| allowed_tools | ["web_fetch"] |
| tags | ["web","extraction","facts","research"] |
Extract Facts
Given web content or a URL, extract a set of discrete, independently verifiable
facts. Each fact must be atomic (one claim per fact), attributed to its source,
and accompanied by a confidence indicator.
Procedure
- Receive source content (raw text, URL, or list of URLs).
- If URLs are provided, fetch each one using
web_fetch.
- Parse the content and identify factual statements. Distinguish:
- Hard facts: specific figures, dates, names, events with a clear source
- Soft facts: general statements of consensus that are widely accepted
- Opinion/editorial: explicitly exclude these
- For each hard fact, record:
- The fact statement (one sentence, active voice)
- The source URL
- The source date
- A confidence level:
high, medium, or low
- Cross-check facts against other sources when possible. Upgrade confidence if
confirmed by a second independent source.
- Return the extracted facts as a structured list.
Output Format
Fact 1:
Statement: <one-sentence factual claim>
Source: <url>
Date: <publication date or "unknown">
Confidence: high | medium | low
Notes: <optional context or caveat>
Fact 2:
...
Quality Rules
- Each fact must be independently verifiable from its cited source.
- Avoid compound facts; split "X happened in Y and caused Z" into two facts.
- Mark speculation or prediction as
Confidence: low with a note.
- If a URL cannot be fetched, record the fact as
Confidence: low with a note
indicating the source was inaccessible.
- Do not include the agent's own inferences as facts; attribute all claims to
an external source.