Fact-Check Pass

---
name: fact-check-pass
description: Reads a finished draft looking only for what could be wrong: numbers with no source, facts that have quietly gone out of date, borrowed statistics, claims stated more strongly than the evidence allows. It returns each one with a specific fix. Use as the last pass before publishing, sending or presenting anything that states facts.
---

# Fact-Check Pass

You are the last reader before the thing goes out. You are not editing for style, structure or tone. You are looking for the sentence that will be wrong in public.

## The rule that keeps you useful

Flag, do not rewrite. The author decides what to change. Your job is to make the risk visible and the fix easy, one claim at a time.

The second rule follows from the first: never confirm a fact you cannot actually check. "Unverified" is an honest and useful finding. A false confirmation is worse than saying nothing, because it ends the checking.

## Workflow

1. **Read once for meaning.** You cannot judge whether a claim is overstated until you know what the piece is arguing.
2. **Ask two questions before checking anything.** Where is this being published, and who reads it? A slide for six colleagues and a press release carry different risk, and risk sets how hard you push.
3. **Extract every checkable claim.** Numbers, dates, names and titles, quotes, superlatives, comparisons, cause-and-effect statements, and anything attributed to "studies" or "research".
4. **Sort each claim** into supported, unverified, out of date, overstated, or wrongly attributed.
5. **Check what can be checked** against the sources the author gave you. If you can browse, verify and say what you verified against. If you cannot, say so plainly and flag rather than guess.
6. **Write one finding per claim**: the quote, what is wrong with it, why in one line, and the exact replacement wording or the source that would fix it.
7. **Give a risk verdict and a must-fix list.** Most drafts have a handful of findings that matter and a long tail that does not. Say which is which.

Copy this checklist into your working notes and tick it as you go:

```text
Fact-check checklist
[ ] Read once for meaning before checking anything
[ ] Audience and venue known (sets the risk bar)
[ ] Every number, date, name, quote and superlative extracted
[ ] Borrowed stats traced to a real origin, not to whoever repeated them
[ ] Anything time-sensitive checked for staleness: prices, versions, headcounts, titles
[ ] Each finding has a specific fix, not "consider revising"
[ ] Nothing silently rewritten
[ ] Risk verdict and a short must-fix list
```

## What counts as a checkable claim

| Kind | What to look for | Usual fix |
|---|---|---|
| Numbers | Any figure, percentage, count, money amount | Add the source, or give the figure a date |
| Dates and timelines | "Since 2019", "last quarter", "recently" | Replace vague time words with the actual date |
| Names and titles | Job titles, company names, product names | Verify the person still holds the role |
| Quotes | Anything in quotation marks | Check wording and who said it, and where |
| Superlatives | First, best, fastest, only, largest, leading | Add the measure and the source, or drop the claim |
| Comparisons | "Twice as fast", "cheaper than" | Name what it is being compared against, and when |
| Cause and effect | "X led to Y", "because of X" | Downgrade to correlation unless the source shows cause |
| Borrowed research | "Studies show", "research suggests" | Name the specific study, or cut the sentence |
| Implied claims | Not stated, but clearly meant | Make the implication explicit and check that instead |

The last row catches what everyone misses. A sentence can be technically true and still leave the reader with a false belief. Check what the reader will conclude, not only what the words say.

## Sorting a claim

- **Supported.** You checked it against a source that actually says it. Note the source; do not list it as a finding.
- **Unverified.** Might well be true, but nothing in the draft or the sources supports it. This is the most common finding and the least dramatic. It is not an accusation.
- **Out of date.** Was true, probably is not now. Prices, headcounts, version numbers, job titles, "the current X", anything ranked.
- **Overstated.** The evidence supports something narrower than what is written. Usually a superlative, an "all" or "every", or a cause claim standing on a correlation.
- **Wrongly attributed.** The number is real but came from somewhere else, or the quote belongs to someone else, or a summary is credited to the study it summarised.

Keep "wrong" and "too strong" apart. They read similarly and they need different fixes: one needs a correction, the other needs narrower wording.

## Tracing a borrowed statistic

Numbers get repeated until they lose their parents. When a draft says "studies show that 70% of...", walk it back:

1. Who does the draft credit? Usually an article, not a study.
2. Who did that article credit? Often another article.
3. Does the original say what everyone claims it says? Frequently it says something narrower, about a different population, in a different year.

If you cannot reach an origin, the finding is "cannot be traced to a source - either name the study or cut the sentence". That is a legitimate result, not a failure.

## Output shape

For each finding:

- **The claim:** quoted exactly from the draft.
- **Issue:** unverified, out of date, overstated, wrongly attributed, or needs a source.
- **Why:** one line.
- **Fix:** the replacement wording, or the specific source to add.

Then:

- **Risk:** low, medium or high, with one sentence of reasoning tied to the audience.
- **Must fix before this goes out:** the short list. Two to five items, in order.
- **Fine as written:** one line naming what you checked and found sound, so the author knows the whole draft was read.

## Example

**Input from the user**

> Final check on our launch post before it goes on the site and to press.
>
> "Since launching in 2019, Meridian has become the fastest reporting tool on the market, trusted by millions of users worldwide. Studies show that 70% of finance teams waste over a day a week on manual reporting. Our new engine cuts reporting time by 90%, which is why enterprise adoption tripled last year."

**Findings**

- **Claim:** "the fastest reporting tool on the market"
  **Issue:** overstated, no measure.
  **Why:** fastest at what, measured how, against which competitors? As written it is an advertising claim that a competitor could challenge.
  **Fix:** "the fastest reporting tool in our own benchmark against X and Y, published here" with a link, or cut it.

- **Claim:** "trusted by millions of users worldwide"
  **Issue:** unverified.
  **Why:** nothing in the material supports the number, and "millions" is the kind of figure a journalist will ask you to substantiate.
  **Fix:** use the real registered-user figure with its date, or "used by teams in over N countries" if that is what you can evidence.

- **Claim:** "Studies show that 70% of finance teams waste over a day a week"
  **Issue:** wrongly attributed, cannot be traced.
  **Why:** "studies show" credits nobody. Traced back, this kind of figure usually comes from one vendor survey with a small self-selected sample.
  **Fix:** name the specific study, its year and its sample, or cut the sentence. If it is your own survey, say so: "in our survey of N finance teams".

- **Claim:** "cuts reporting time by 90%"
  **Issue:** needs a source and a baseline.
  **Why:** 90% against what starting point, for which customer, doing what?
  **Fix:** "cut reporting time from around six hours to under forty minutes for [named customer], in their own measurement".

- **Claim:** "which is why enterprise adoption tripled last year"
  **Issue:** overstated cause.
  **Why:** "which is why" claims the engine caused the growth. The engine is new; the growth is from last year. The timeline does not support the causal link.
  **Fix:** "enterprise adoption tripled last year" as a separate sentence, with the figures behind "tripled".

- **Claim:** "Since launching in 2019"
  **Issue:** check.
  **Why:** founding dates and launch dates are often different, and the press will use whichever you print.
  **Fix:** confirm whether 2019 is the company founding, the product launch, or the first paying customer, and say which.

**Risk: high.** This is going to press, and three of the six claims are the kind a journalist checks first: a superlative, an unnamed study, and a causal claim.

**Must fix before this goes out:** the unnamed study, "millions of users", the 90% baseline, and the causal "which is why".

**Fine as written:** the product description and the pricing paragraph make no factual claims that need support.

## Rules

- Flag, do not silently rewrite. Suggest wording; the author chooses.
- Never invent a source to rescue a claim you like.
- Every finding gets a specific fix. "Consider revising" is not a finding, it is a shrug.
- Say what you checked against. A finding with no basis is just an opinion about the sentence.
- Watch the small print: prices, plan names, job titles, version numbers and anything with "current" in it go stale faster than anything else.
- Do not fact-check opinion. "We think this is the best way to work" needs no source. "This is the best way to work" does.
- Respect the author's voice. You are not here to make the piece more cautious, only more true.

## Edge cases

- **You cannot browse.** Say so in the first line. Flag everything checkable and mark it "needs verification", and be explicit that absence of a flag is not a confirmation.
- **The draft is mostly opinion.** Say so, check the handful of factual claims propping the argument up, and stop. A long list of findings on an opinion piece means you were checking the wrong thing.
- **The author is the source.** "Our customers save an average of four hours." Ask how it was measured, over what sample, and when. Internal numbers go stale and get repeated for years.
- **The claim is legally sensitive:** health, money, safety, or anything about a named competitor. Flag it as needing review by someone qualified, and do not soften it into acceptability yourself.
- **Everything is wrong.** If more than about a third of the claims fail, stop listing and say the piece needs a source pass before a fact check. Forty findings will not be read.
- **The number is right but misleading.** Percentages with no base, growth from a tiny starting point, an average hiding a wide spread. Flag it as misleading and give the honest framing.
- **The draft cites something you cannot access.** Note that it is uncheckable from here and name what the author should confirm in it. Do not treat a paywall as a failure.

## Files

- `references/examples.md` - three drafts checked end to end: a company announcement, a social post built on a borrowed statistic, and an internal board update with a stale number. Read it to see how findings are worded and how the risk verdict is reasoned.
- `references/claim-types.md` - the full claim taxonomy with the question to ask for each type, the standard fix, and the phrases that reliably hide an unsupported claim. Read it while extracting claims from an unfamiliar kind of document.

# Worked fact-check passes

Three drafts read the way this skill reads them. Each shows the extraction, the findings, and the verdict.

## Contents

- [Example 1: a customer case study](#example-1-a-customer-case-study)
- [Example 2: a social post built on a borrowed statistic](#example-2-a-social-post-built-on-a-borrowed-statistic)
- [Example 3: an internal board update](#example-3-an-internal-board-update)
- [What the three have in common](#what-the-three-have-in-common)

---

## Example 1: a customer case study

### The draft

> Northwind Logistics had been struggling with manual dispatch for over a decade. After switching to our platform in Q2, their dispatch time fell by half and driver satisfaction rose sharply. "This changed everything for us," says their operations director. Northwind now processes 40% more deliveries with the same headcount, and they are one of the largest independent hauliers in the region.

**Venue:** the company website, linked from sales emails. **Reader:** prospective buyers, and Northwind themselves.

That second reader matters. A case study is checked hardest by the customer it describes, and an overstatement damages that relationship before it damages anything else.

### Findings

- **Claim:** "struggling with manual dispatch for over a decade"
  **Issue:** unverified, and it characterises the customer.
  **Why:** "struggling for a decade" is a judgement about Northwind that Northwind has to be comfortable seeing in public.
  **Fix:** use their own wording from the interview, or "had run dispatch manually since [year]".

- **Claim:** "dispatch time fell by half"
  **Issue:** needs a baseline and a measurement window.
  **Why:** half of what, measured over how long, and by whom? A single good week is not a result.
  **Fix:** "average dispatch time fell from X minutes to Y minutes across the first three months, measured by Northwind".

- **Claim:** "driver satisfaction rose sharply"
  **Issue:** unverified.
  **Why:** satisfaction implies a survey. If there was no survey, this is somebody's impression.
  **Fix:** cite the survey and the change, or replace with a quote from a driver, or cut.

- **Claim:** "This changed everything for us"
  **Issue:** check attribution and approval.
  **Why:** quotes get tidied in drafting, and the named person has to stand behind the final wording.
  **Fix:** confirm the exact sentence and that the operations director has approved it by name and title.

- **Claim:** "40% more deliveries with the same headcount"
  **Issue:** needs a period, and possibly misleading.
  **Why:** if the 40% covers a seasonal peak, the comparison is against a quieter period and the figure flatters itself.
  **Fix:** state the two periods compared, and prefer a year-on-year comparison.

- **Claim:** "one of the largest independent hauliers in the region"
  **Issue:** overstated, no measure.
  **Why:** largest by what, in which region, according to whom.
  **Fix:** cut it, or use a factual descriptor: "operates N vehicles across [area]".

- **Claim:** "switching to our platform in Q2"
  **Issue:** out of date on arrival.
  **Why:** "Q2" with no year stops making sense in six months and the page will not be revisited.
  **Fix:** name the month and year.

**Risk: medium.** Nothing here is likely to be false, but four claims cannot be defended if Northwind's own team reads them closely, and they will.

**Must fix before this goes out:** the dispatch-time baseline, the 40% period, quote approval, and the "largest independent" claim.

**Fine as written:** the description of what the platform does and the implementation timeline.

**Why it is shaped this way:** the venue analysis changed the whole pass. Because the subject of the piece is also a reader, claims that characterise them get flagged even when they are probably accurate.

---

## Example 2: a social post built on a borrowed statistic

### The draft

> 90% of startups fail. But here is what nobody tells you: the ones that survive all do the same three things. Research shows that companies with a written strategy grow 30% faster. That is not a coincidence. If you are not writing your strategy down, you are already losing.

**Venue:** a personal professional post. **Reader:** peers and potential clients.

### Findings

- **Claim:** "90% of startups fail"
  **Issue:** wrongly attributed, cannot be traced.
  **Why:** this figure circulates constantly with no consistent origin. Different sources define "startup" and "fail" differently, and over different time windows, which is why the number ranges enormously depending on who is quoting it.
  **Fix:** name a specific source with its definition and window, or replace with the honest version: "most new businesses do not last, and the exact rate depends on how you count".

- **Claim:** "the ones that survive all do the same three things"
  **Issue:** overstated, and it is survivorship reasoning.
  **Why:** "all" is doing heavy lifting. Looking only at survivors cannot tell you what caused survival, because the failures may have done the same three things.
  **Fix:** "three habits I see repeatedly in the ones that last", a claim about your observation, which you can support.

- **Claim:** "Research shows that companies with a written strategy grow 30% faster"
  **Issue:** wrongly attributed, and likely a cause claim resting on a correlation.
  **Why:** "research shows" names nobody. Even if a study found the association, organised companies both write strategies and grow, so the writing may be a symptom rather than a cause.
  **Fix:** name the study, its year and sample, and state it as an association: "companies with a written strategy grew faster in [study]; that does not prove the document caused it".

- **Claim:** "That is not a coincidence"
  **Issue:** overstated cause.
  **Why:** this sentence converts a correlation into a cause in four words.
  **Fix:** cut it, or replace with what you actually believe and why.

- **Claim:** "you are already losing"
  **Issue:** not a factual claim.
  **Why:** rhetoric, not fact. Out of scope for this pass.
  **Fix:** none needed.

**Risk: medium.** Nothing here is dangerous, but the post's whole structure rests on two untraceable numbers, and the audience is exactly the audience that checks.

**Must fix before this goes out:** the 90% figure and the "research shows" sentence. Fixing those two forces the argument to stand on the author's own experience, which is where it was always strongest.

**Fine as written:** the closing call to action makes no factual claim.

**Why it is shaped this way:** the last finding is a deliberate non-finding. Marking rhetoric as out of scope shows the author that you can tell the difference, which is what makes them trust the findings you did raise.

---

## Example 3: an internal board update

### The draft

> Churn is down to 3.1% monthly, our best ever. The support team of twelve is now handling 40% more tickets since Ravi took over as Head of Support. Cash runway currently sits at 14 months. Our largest competitor recently raised prices, which is driving the inbound we are seeing.

**Venue:** a board pack. **Reader:** investors and non-executive directors who will quote these numbers elsewhere.

### Findings

- **Claim:** "Churn is down to 3.1% monthly, our best ever"
  **Issue:** needs a definition and possibly misleading.
  **Why:** customer churn or revenue churn? They differ, sometimes a lot, and boards assume revenue. "Best ever" also needs the comparison period stated.
  **Fix:** "monthly revenue churn of 3.1% in [month], against 4.0% the same month last year" and say which measure.

- **Claim:** "the support team of twelve"
  **Issue:** out of date risk.
  **Why:** headcount is the fastest-staling number in any board pack, and it will be read weeks after it is written.
  **Fix:** "twelve as of [date]".

- **Claim:** "40% more tickets since Ravi took over"
  **Issue:** overstated cause.
  **Why:** "since" invites the reader to hear "because of". More tickets may also mean more customers, or a product problem generating them. Rising ticket volume is not obviously good news.
  **Fix:** separate the two: state the volume change with its period, and state tickets per customer, which is the number that tells you whether anything improved.

- **Claim:** "Head of Support"
  **Issue:** check the title.
  **Why:** titles in board packs get quoted back in introductions and press.
  **Fix:** confirm the exact title against the current org chart.

- **Claim:** "Cash runway currently sits at 14 months"
  **Issue:** needs its assumptions.
  **Why:** runway depends entirely on assumed burn. At current burn, at planned burn including the hires in this same pack, and at flat revenue, that number can differ by months.
  **Fix:** "14 months at current burn; 11 months including the two hires proposed in section 4".

- **Claim:** "Our largest competitor recently raised prices, which is driving the inbound"
  **Issue:** unverified and overstated cause.
  **Why:** two claims in one sentence. The price rise is checkable. The causal link to inbound is not, unless inbound leads were asked.
  **Fix:** state the price change with its date and source, then state inbound separately, with whatever attribution data exists.

**Risk: high.** Not because anything is likely false, but because board numbers travel. A director will repeat "3.1% churn" and "14 months runway" in a room where nobody knows the assumptions.

**Must fix before this goes out:** the churn definition, the runway assumptions, and splitting the two claims in the competitor sentence.

**Fine as written:** the product roadmap section states plans, not facts, and needs no sourcing.

**Why it is shaped this way:** the risk verdict is high on a document with no false statements in it. Risk is about consequence and travel, not just accuracy, and internal documents are where that distinction earns its keep.

---

## What the three have in common

- The venue and the reader are established first, and they change what gets flagged. The same sentence is fine in a Slack message and a problem in a press release.
- Every finding names a specific fix, in the author's voice, ready to paste.
- Cause claims hiding inside connecting words ("since", "which is why", "that is not a coincidence") are flagged in all three. They are the most common way a draft claims more than it knows.
- Findings that are out of scope are said out loud rather than silently skipped, so the author can see the whole draft was read.
- The must-fix list is short. A pass that flags everything equally has not done the reader's judging for them.

# Claim types and their fixes

A lookup for the extraction step. Find the kind of claim, ask its question, apply its fix.

## Contents

- [Numbers](#numbers)
- [Dates and time words](#dates-and-time-words)
- [Names, titles and organisations](#names-titles-and-organisations)
- [Quotes](#quotes)
- [Superlatives and rankings](#superlatives-and-rankings)
- [Comparisons](#comparisons)
- [Cause and effect](#cause-and-effect)
- [Borrowed research](#borrowed-research)
- [Implied claims](#implied-claims)
- [Phrases that hide an unsupported claim](#phrases-that-hide-an-unsupported-claim)
- [Numbers that are true and still misleading](#numbers-that-are-true-and-still-misleading)
- [What not to fact-check](#what-not-to-fact-check)

---

## Numbers

**Ask:** where did this come from, when was it measured, and what is the base?
**Fix:** attach the source and the date. If it is an internal figure, say so and name the period.
**Watch:** round numbers that have been repeated so often nobody remembers measuring them.

## Dates and time words

**Ask:** does this still mean what it meant when it was written?
**Fix:** replace relative time with absolute. "Recently", "currently", "last quarter", "this year" and "new" all rot.
**Watch:** documents that will be read months after they are written: web pages, case studies, help articles.

## Names, titles and organisations

**Ask:** does this person still hold this role, and is the organisation still called this?
**Fix:** verify against a current source, and add the date if the title is load-bearing.
**Watch:** anything copied from an older document. Titles are the most-copied stale fact there is.

## Quotes

**Ask:** are these the exact words, is this the right speaker, and have they seen the final wording?
**Fix:** confirm the sentence verbatim and get approval where the quote is attributed by name.
**Watch:** quotes tidied during drafting. A cleaned-up quote is a misquote, however small the improvement.

## Superlatives and rankings

**Ask:** best at what, measured how, compared with whom, according to which source?
**Fix:** add the measure and the source, narrow the claim, or cut it.
**Watch:** first, best, fastest, only, largest, leading, top-rated, industry-standard, award-winning.

## Comparisons

**Ask:** compared with what, at what moment, under what conditions?
**Fix:** name the comparison point and the date. "Twice as fast as our previous version, in our own testing on [date]".
**Watch:** comparisons against an unnamed "the alternative" or "traditional approaches", which cannot be checked or challenged.

## Cause and effect

**Ask:** does the source show that X caused Y, or only that both happened?
**Fix:** downgrade to association and say so. "Adoption tripled in the same period" instead of "which is why adoption tripled".
**Watch:** the causal link is usually in a small connecting word rather than in the claim itself. See the phrase list below.

## Borrowed research

**Ask:** which study, what year, what sample, and does it say what the draft says it says?
**Fix:** name study, year and sample size, or cut the sentence.
**Watch:** the three-step chain, where the draft cites an article, the article cites another article, and the original study says something narrower about a different group.

## Implied claims

**Ask:** what will the reader believe after reading this, and is that supported?
**Fix:** make the implication explicit, then check that sentence instead.
**Watch:** two true sentences placed next to each other so the reader infers a link that neither one states. This is the hardest category to spot and the one that causes real trouble.

## Phrases that hide an unsupported claim

| Phrase | What it is doing |
|---|---|
| "Studies show" / "research suggests" | Citing nobody |
| "It is well known that" | Citing nobody, more confidently |
| "Which is why" / "that is why" | Turning a correlation into a cause |
| "Since [event]" | Inviting a causal reading without claiming one |
| "Up to" | Reporting the best case as if it were typical |
| "Industry-leading" / "best-in-class" | A ranking with no ranker |
| "Trusted by thousands" | A number with no measurement |
| "Proven to" | A claim of evidence that is rarely attached |
| "Experts agree" | Averaging unnamed people |
| "That is not a coincidence" | Asserting cause in place of evidence |
| "Everyone knows" / "nobody talks about" | Claiming consensus in either direction |

## Numbers that are true and still misleading

- **A percentage with no base.** "Up 300%" from three customers to twelve is technically accurate and practically meaningless without the base.
- **An average hiding a spread.** "Customers save an average of four hours" where two save twenty and the rest save none.
- **A cherry-picked window.** A comparison starting at the worst month in the record.
- **A different denominator.** Percentage of users against percentage of active users against percentage of paying users.
- **Percent against percentage point.** Moving from 2% to 3% is one percentage point and a 50% increase. Both are true; one is chosen.
- **A total presented as a rate**, or a rate presented as a total.
- **A figure not adjusted for inflation** in any statement about money across years.

Each of these is a finding worded as "true but misleading", with the honest framing supplied.

## What not to fact-check

- Opinions clearly marked as opinions. "We think this is the better approach" needs no source.
- Predictions about the future, unless they are stated as certainties.
- Rhetoric and calls to action that make no factual claim.
- Style, structure and tone. Another pass owns those, and mixing them in buries the findings that matter.
- Claims the author has already sourced correctly. Note them once under "fine as written" and move on.