/claim-check

---
name: claim-check
description: >-
  Before anything is published or sent, extract every checkable claim in a
  draft (statistics, dates, prices, quotes, names, job titles, "first",
  "largest", "studies show") and verify each against primary sources on the
  web, returning a claim-by-claim verdict with links and a corrected
  version. Use this skill whenever the user says "fact-check", "is this
  true", "verify this", "check before I post", "can I say this", or is about
  to publish a LinkedIn post, newsletter, article, slide, press release or
  client email that contains facts. ALSO use it proactively before
  delivering any public-facing text with numbers or quotes in it. Do NOT use
  to check numbers against the user's own files (use number-check), for
  opinion pieces, or for private brainstorming.
---

# Claim check

A post with one wrong statistic loses more than that statistic. The reader
who catches it, often in the comments, stops trusting everything else you
wrote, and the correction never travels as far as the original. On
LinkedIn, the most shared numbers are often the least checked: a figure
quoted by a hundred accounts usually traces back to one source, sometimes
to none.

AI drafts make this worse in a specific way: they produce numbers that
sound right. A plausible statistic with a plausible source attached is the
hardest kind of error to spot. So every claim gets checked on purpose, one
by one, against the place it should come from.

## Workflow

### 1. Extract the claims

Number every checkable statement and keep its exact wording:
"the largest" and "one of the largest" are different claims. Hidden
claims count: "doubled", "most", "record", "since 2020", "according to
Harvard". Skip opinions, predictions and the author's own experiences.

### 2. Check the riskiest first

1. **Numbers:** statistics, percentages, prices, dates, counts.
2. **Quotes:** exact words AND who said them. Misattribution is common.
3. **Names and titles:** spelling, current role, company.
4. **Superlatives:** first, only, largest, never. One counterexample kills them.
5. **Money, health and legal claims:** primary sources only.

### 3. Verify at the source

- Go to the origin: the study, the official statistics office, the
  company filing, the transcript. A news article quoting a study is a
  pointer, not proof: follow it.
- Check the date and the scope: a 2022 survey of US managers does not
  prove a 2026 claim about "professionals".
- Surprising claims need two independent sources that do not cite each other.
- No findable origin is a finding in itself.

### 4. One verdict per claim

- **TRUE:** matches a reliable source. Link it.
- **OUT OF DATE:** was true; give the current figure and its date.
- **STRETCHED:** right direction, wrong as written (rounded up, "proves"
  where the source says "linked to", "studies" where there is one study).
  Give honest wording.
- **NO SOURCE:** cannot be traced. Cut it or hedge it.
- **FALSE:** contradicted by a reliable source. Show what it says.

### 5. Report and repair

```
## Claim check: N claims. X true, Y to fix, Z no source.

1. "exact claim": STRETCHED
   Source: [name, link, date] says [what it actually says]
   Use instead: "[corrected wording]"
2. ...
```

Put problems first. Then offer the corrected draft, changing only what the
verdicts require. Do not restyle the writing at the same time.

## Rules

**No link, no verdict.** Every TRUE and every FALSE carries a source the
user can click.

**Never check from memory.** If web search is unavailable, say so at the
top, list the claims with your confidence, and mark everything unverified.

**A clean sheet must be earned.** If every claim passes in a stat-heavy
draft, re-check the two most surprising ones.

**Overclaiming is an error.** "Proven" when the source says "associated"
is STRETCHED, even if the number is right.

## Example

Draft: "Studies show 8 in 10 workers now use AI every day."
One sentence, three checks. "Studies" (plural): the search finds one
survey, so STRETCHED. "8 in 10": the survey reports 78% of respondents who
already use AI tools at work, a narrower group than "workers", so
STRETCHED again. "Every day": the survey asked about weekly use, so FALSE
as written. Use instead: "In one 2026 survey, about 3 in 4 people who use
AI at work said they use it weekly." The claim got smaller and became
defensible, which is the trade every good post makes.

(The survey in this example is illustrative. When you run the skill, every
verdict must point to a real, linked source.)