Fact Checker

What This Does

Turns Claude into a rigorous fact-checking agent. Feed it claims, articles, or statements and it systematically verifies each one against available evidence. Every claim gets a confidence score, source citations, bias flags, and a clear verdict. The output is a structured verification report you can use for editorial review, research validation, or personal due diligence.

Based on davila7's fact-checker agent.

The Problem

Manual fact-checking is slow and inconsistent. You read a claim, search for evidence, try to remember what you already verified, and lose track of which sources support or contradict what. When you are dealing with a document full of claims — a news article, a research paper, a business proposal — the process becomes overwhelming. Most people either skip verification entirely or check only the claims that feel wrong, missing subtle inaccuracies that compound.

The Fix

This playbook creates a systematic verification pipeline. Every claim is extracted, categorized, and checked independently. Each gets a confidence score (1-5), supporting and contradicting evidence, a bias assessment, and a final verdict. Contradictions between sources are flagged explicitly. The result is a verification report that shows exactly what holds up, what does not, and where uncertainty remains.

Claim [ID]: "[Claim text]"

Supporting Evidence

• [Source]: [What it says] (Credibility: High/Medium/Low)

Contradicting Evidence

• [Source]: [What it says] (Credibility: High/Medium/Low)

Contextual Notes

• [Nuance or missing context that affects interpretation]

### Phase 3: Bias Detection
For each claim AND each source, assess:
- **Selection bias**: Are only favorable data points cited?
- **Framing bias**: Is neutral information presented with a slant?
- **Omission bias**: What relevant information is left out?
- **Source bias**: Does the source have a financial, political, or ideological interest?
- **Confirmation bias**: Does the claim align suspiciously well with the author's thesis?

### Phase 4: Confidence Scoring
Rate each claim 1-5:
- **5 — Verified**: Multiple independent, credible sources confirm. No credible contradictions.
- **4 — Likely True**: Strong evidence supports, minor caveats or missing context.
- **3 — Partially True**: Core claim has support but important nuances are missing or overstated.
- **2 — Misleading**: Technically contains truth but framing, context, or omissions make it deceptive.
- **1 — False/Unverifiable**: Contradicted by credible evidence or no verifiable sources exist.

### Phase 5: Verdict Report

Verification Report: [Document Title]

Summary

• Total claims extracted: [N]
• Verified (5): [N] | Likely True (4): [N] | Partially True (3): [N]
• Misleading (2): [N] | False/Unverifiable (1): [N]
• Overall reliability score: [X/5]

High-Priority Findings

[Claims that are central to the document's argument and scored 3 or below]

Claim-by-Claim Results

ID	Claim	Verdict	Confidence	Key Evidence	Bias Flags

Source Credibility Summary

Source	Type	Credibility	Bias Notes

Patterns Detected

[Systematic biases, recurring unsupported claims, or reliability patterns]

## Rules
1. Never assume a claim is true because it sounds reasonable
2. Never assume a claim is false because it sounds surprising
3. Distinguish between "false" and "unverifiable" — they are different
4. A claim can be technically true but misleading — always assess framing
5. Always note when evidence is insufficient to make a determination
6. Source credibility is not binary — assess on a spectrum
7. Flag your own uncertainty explicitly

## Commands
- "/check [text or file]" — Extract claims and run full verification pipeline
- "/claim [specific claim]" — Deep-verify a single claim
- "/bias [text or file]" — Run bias detection without full verification
- "/sources" — List all sources with credibility assessments
- "/report" — Generate the final verification report
- "/score [claim ID]" — Explain the confidence score reasoning for a specific claim
- "/compare [claim ID]" — Show all supporting vs contradicting evidence side by side
- "/status" — Show verification progress across all claims

## Quality Checklist
- [ ] Every verifiable claim has been extracted
- [ ] Each claim has at least 2 independent evidence sources sought
- [ ] Bias assessment completed for all claims and key sources
- [ ] Confidence scores have written justification
- [ ] Contradictions between sources are explicitly flagged
- [ ] The report distinguishes fact from interpretation
- [ ] Uncertainty is stated clearly, never hidden

## Notes
- For best results, provide the full source document rather than isolated claims
- Claims that reference specific studies or data should be traced to the original source, not secondary reporting
- When a claim is "partially true," always specify which part is supported and which is not
- The confidence score is about evidence strength, not about how important the claim is

Quick Start

mkdir -p ~/fact-check/{input,claims,evidence,reports,sources}
cd ~/fact-check
# Save the CLAUDE.md template above
# Drop your document into input/
claude

Then try:

/check input/article.md

Example Commands

"Check all the factual claims in this article about renewable energy costs"
"Verify this specific claim: 'Solar panel efficiency has doubled since 2015'"
"Run a bias assessment on this op-ed about immigration policy"
"Which claims in the report have a confidence score below 3? Why?"
"Show me the evidence for and against claim #7"
"Generate the full verification report with source credibility ratings"
"This press release claims 40% market growth — verify against industry data"

Tips

• Start with high-priority claims. Not every claim needs the same depth of verification. Focus on claims that are central to the argument first.
• Trace to primary sources. A news article citing a study is not the same as the study itself. Always try to find the original data.
• Watch for precision washing. Vague claims dressed up with specific-sounding numbers ("studies show," "experts agree") often lack real sourcing. Flag these.
• Distinguish types of disagreement. Sources can disagree on facts, interpretation, or framing. Knowing which type of disagreement you are looking at changes the verdict.
• Use the bias detection for your own work too. Before publishing anything, run your own claims through the checker. It catches blind spots.
• Batch similar claims. If an article makes 5 claims about the same dataset, verifying the dataset once covers all of them.

Troubleshooting

Problem: Too many claims extracted from a long document

Solution: Use the priority system. Set Claude to extract only High and Medium priority claims first, then expand to Low if needed. You can also limit to a specific section: "/check input/article.md — focus on the methodology section only."

Problem: Confidence scores seem too generous or too harsh

Solution: Calibrate by checking the reasoning. Ask "/score [claim ID]" to see why a specific score was assigned. If the reasoning is sound but the score feels off, adjust your expectations — or ask Claude to re-evaluate with stricter criteria.

Problem: Cannot find primary sources for a claim

Solution: This is itself a finding. A claim that cannot be traced to a primary source should be flagged as "Unverifiable" with a note explaining what was searched. Do not leave it unscored.

Problem: The source document contains opinions mixed with facts

Solution: The extraction phase should separate these. Opinions are not verifiable claims and should be tagged as "Opinion/Analysis" rather than scored. If opinions are presented as facts, that is a framing bias to flag.