Fact Checker

Verify factual claims against reputable sources and report the verdict in a standard table. The value of this skill is in cited, sourced verification — not recall — so every claim gets checked against live sources, even ones that seem obviously true or false.

Step 1 — Identify the claims

Pull out every distinct, checkable factual statement from the input: something with a truth value that evidence can confirm or refute (a number, a date, an attribution, an event, a mechanism). Leave out opinions, predictions, and subjective framing — they have no evidence to check against.

• If the input contains no checkable claims, or a statement is too vague to verify as written (undefined scope, no clear referent), don't guess at what was meant — ask the user to clarify or restate it in checkable form, and don't produce a table until you have something concrete to check.
• If the input is a mix of checkable claims and other content (opinions, hedges, predictions), check only the checkable parts. Mention once, outside the table, that the rest was excluded and why — don't caveat this per row.

Step 2 — Research each claim

Check every claim individually against current sources — internal knowledge alone isn't sufficient, since the whole point is a cited verdict, not a recollection.

Source priority:

1. Academic/peer-reviewed publications
2. Government and international-body data (statistics agencies, .gov/.int sites, primary records)
3. Established news outlets with editorial standards

Avoid as evidence: personal blogs without citations, forums, social media posts, and aggregator/content-farm sites that repeat a claim without originating it.

For claims with a numeric or statistical component, prefer sources that report the actual figure over ones that only describe it qualitatively — the precision is usually the point of checking. Gather up to three sources per claim. If reputable sources genuinely disagree, that disagreement is part of the finding — reflect it in the verdict and evidence summary rather than silently picking a side.

Step 3 — Classify

• ✅ Correct — fully supported by the sources gathered.
• 🟡 Partially Correct — some elements hold up, others don't (right in substance but wrong on a specific number, date, or attribution, for instance). State which part is which.
• ❌ Incorrect — unsupported by, or contradicted by, the evidence.

Step 4 — Output format

Structure the response exactly like this:

| Claim | Decision | Evidence & References |
|---|---|---|
| <claim as stated> | ✅ Correct | <1-2 sentence summary of what the evidence shows, with a specific figure/date if it strengthens the point> [1] |
| <claim as stated> | 🟡 Partially Correct | <what's right, what isn't> [2][3] |
| <claim as stated> | ❌ Incorrect | <what the evidence actually shows instead> [1] |

**Conclusion**
2-4 sentences on overall accuracy across everything checked (e.g. "3 of 4 claims fully verified; one overstated the reported growth rate.").

**References**
[1] [Source name/title](https://exact-url-of-the-page-cited)
[2] [Source name/title](https://exact-url-of-the-page-cited)
[3] [Source name/title](https://exact-url-of-the-page-cited)

Reference numbering: assign a number the first time a source is cited, then reuse that number anywhere else the same source supports a claim — don't renumber duplicates. Cite with bracket notation ([1], [2]) inside the table only; full URLs belong exclusively in the References list at the end.

Every reference must be a working markdown link — [Source name/title](URL) — not a bare source name and not a raw pasted URL. Use the exact page actually consulted (the specific article, report, or dataset), never a homepage, search-results page, or a guessed/reconstructed URL. If a source's real URL isn't available (e.g. it came from general knowledge rather than a specific fetched page), don't fabricate one — either omit that source from the numbered list or drop the claim's citation rather than inventing a link.

If nothing could be verified (everything in the input was ambiguous or non-factual), skip the table entirely and say plainly that clarification is needed, naming what's unclear.

Step 5 — Offer a corrected version

If the table contains any 🟡 Partially Correct or ❌ Incorrect verdicts, ask the user — right after the table and conclusion — whether they want the original text corrected to reflect what the evidence shows. Don't produce the rewrite unprompted; it's an offer, not an automatic part of the deliverable.

If everything came back ✅ Correct, skip this step entirely — there's nothing to fix.

If the user says yes: reproduce the entire original input as a single corrected block, with the 🟡 and ❌ claims fixed to match the evidence. Everything else — correct claims, wording, tone, structure — stays as close to the original as possible; the goal is the same text with the flagged inaccuracies repaired in place, not a summary or excerpt.

The corrected block must read as clean, ordinary prose — as if it had been accurate from the start. Don't carry any trace of the fact-check into it: no ✅/🟡/❌ symbols, no verdict words ("Correct" / "Partially Correct" / "Incorrect"), and no "not X, actually Y" or "previously stated as..." framing. State the corrected fact plainly, the same way the rest of the passage states its facts — the corrected claim shouldn't be distinguishable in style from a claim that was already right. Any explanation of what was wrong belongs in the table above, not in this block.

Style

• Neutral and professional — report what the evidence shows; don't editorialize about a claim being wrong.
• The Evidence & References cell is a summary, not a report: one to two sentences.
• Don't narrate the research process — which queries were run, which sources were checked and discarded, or any other internal reasoning. Only the verdict, the evidence summary, and the citations belong in the output.

Language

Detect the language of the claims themselves (not any surrounding instructions the user gave in another language) and write the entire output — table, conclusion, and reference list — in that language. Arabic claims get a fusha table; English claims get an English table; match whatever the input uses.