Skip to main content
ClaudeWave
Skill70 estrellas del repoactualizado 2mo ago

fact-check-before-trust

fact-check-before-trust is a verification skill that validates factual accuracy in agent outputs by extracting claims, scoring confidence levels, and cross-referencing sources before the output is used. Use it when outputs contain numbers, dates, named entities, causal claims, or superlatives in research, financial, legal, or compliance workflows where accuracy directly impacts decisions.

Instalar en Claude Code
Copiar
git clone --depth 1 https://github.com/ArchieIndian/openclaw-superpowers /tmp/fact-check-before-trust && cp -r /tmp/fact-check-before-trust/skills/core/fact-check-before-trust ~/.claude/skills/fact-check-before-trust
Después abre una sesión nueva de Claude Code; el skill carga automáticamente.

SKILL.md

# fact-check-before-trust

`verification-before-completion` checks that a task was *done*. This skill checks that the *facts are correct*. An agent confidently reporting a £716 visa fee as £70,000 will pass completion verification — this skill catches it.

## When to invoke

Invoke this skill before treating any agent output as authoritative when the output contains:
- **Numbers or money** (prices, quantities, measurements, statistics)
- **Dates and deadlines** (filing deadlines, release dates, expiry dates)
- **Named entities** (people, organisations, laws, product names)
- **Causal claims** ("X causes Y", "because of Z")
- **Superlatives** ("the largest", "the only", "the most recent")

Skip for: code output, file system operations, and clearly self-contained tasks (renaming a variable, formatting a document).

## Verification protocol

**Step 1 — Extract claims**
Identify every verifiable claim in the output. List them explicitly:
```
Claim 1: UK visa fee is £716
Claim 2: Processing time is 3 weeks
Claim 3: Applies to Tier 2 (Skilled Worker) visa category
```

**Step 2 — Score each claim**
For each claim, assign a confidence level:
- **High** — Agent has direct evidence in its context (read a document, fetched a URL)
- **Medium** — Agent inferred from training data (check recency)
- **Low** — Agent stated without citing a source

**Step 3 — Verify low/medium claims**
For each Low or Medium claim:
1. Search or re-fetch the source if possible
2. If source found: update confidence to High or mark **Contradicted**
3. If no source available: mark **Unverifiable**

**Step 4 — Classify output**
| Result | Meaning |
|---|---|
| ✓ Verified | All claims High confidence |
| ⚠ Uncertain | 1+ Unverifiable claims |
| ✗ Contradicted | 1+ claims conflict with found evidence |

**Step 5 — Surface to user**
- **Verified**: proceed
- **Uncertain**: surface unverifiable claims with a note
- **Contradicted**: stop, show the contradiction, do not use the output until resolved

## Difference from verification-before-completion

`verification-before-completion` checks: "Did the agent do the task?" (task completion)
`fact-check-before-trust` checks: "Is what the agent said true?" (output accuracy)

Both should be used for research, financial, legal, and compliance workflows.