Review skill for your AI agent
Flag literacy gaps in claims and arguments.
Review skill for your AI agent
Flag literacy gaps in claims and arguments.
Does this claim hold up on basic literacy grounds? ClaimCheck gives your AI agent a seven-dimension review for an article, post, or short argument. You get a report that quotes the claim, says which dimensions fail and why, and names the primary source to check next.
You receive a Markdown report. The source stays unchanged. Anything your agent could not check is marked unverified, not filled in.
A common sentiment about AI and water, and the kind of assertion that merits a more careful check.
1 · The claim Big number, no baseline
These AI data centers are wasting billions of gallons of water a year — water that is gone forever, boiled off so a machine can write someone’s homework or a bad poem.
2 · ClaimCheck’s report Two of seven findings
Excerpt from data-center-water.claimcheck.md
1 · Mechanistic literacy — clear failure
Evaporative cooling does not boil or destroy water. It evaporates part of the circulating water to carry heat away; the vapor re-enters the hydrologic cycle. That is consumption from the local supply, not destruction.
2 · Quantitative literacy — clear failure
LBNL: U.S. data centers directly consumed about 66 billion liters (≈ 17.4 billion gallons) in 2023. USGS: U.S. irrigation consumed 73.2 billion gallons per day in 2015. Share ≈ 0.07%.
3 · What survives The local question
Boundary — partial. “Out where I live the reservoir is low, and a new data center just went in.” That is a real, checkable concern. Next step: the facility’s water permit, its cooling design, and the reservoir’s drought data.
Score 0.5 / 7. The report says which parts fail, which concern survives, and what to read to settle it.
Sample report on an illustrative source, not a recorded tool run. Every figure is dated and sourced in the full report.
Read the complete sample report →What it reads, writes, and changes
| Reads | An article, post, assertion, thread, or short argument. A file, a URL, or pasted text |
| Writes | <name>.claimcheck.md next to a file, or claimcheck.md for a URL or pasted text. Chat only, if you ask |
| Changes | Nothing. The source is never edited |
| Status | Means | Points |
|---|---|---|
| clear failure | The claim commits the dimension’s failure, and it changes the conclusion | 0 |
| partial | Half right, or right in one place and wrong in another | 0.5 |
| passed | A reader could judge the claim on this axis | 1 |
| N/A | The dimension does not apply | left out |
The score is a summary line. Two claims at 3.5 can fail for different reasons, so the findings come first.
The seven dimensions
- Mechanistic literacy. Does the claim correctly model the physical/chemical/operational process? Fails when withdrawal, consumption, recirculation, and destruction are treated as identical.
- Quantitative and comparative literacy. Does it locate magnitude relative to relevant baselines? Fails on absolute numbers without denominators, or selective focus that ignores larger uses.
- Scale and boundary literacy. Does it distinguish local from national or global, direct from indirect? Fails when local stress is flattened into national totals, or the reverse.
- Trade-off and substitution literacy. Does it recognize that reducing one input increases another? Fails when one resource is treated as independently bad, with no substitute named.
- Output and value accounting. Does it count what the resource use produces? Fails when use is called “waste” only because the product is non-physical.
- Recency and updating literacy. Does it reflect current data, technology, and revisions? Fails on superseded estimates.
- Source and evidentiary literacy. Does it prefer primary measurements over secondary amplification? Fails when a single comparison is promoted over systematic reports.
Same seven questions, other subjects (illustrative)
Water is one example. The dimensions are worded for resource claims, and they read the same way on a health risk or a public statistic: what the number measures, what the baseline is, where the boundary sits, what the swap costs. The subjects below differ on purpose. ClaimCheck asks the same questions whichever way a claim leans.
A health-risk claim
A common sentiment about processed meat and cancer.
# ClaimCheck — "Bacon is the new cigarettes"
**Summary:** clear failures: 2 · partial: 4 · passed: 0 · N/A: 1
**Score:** 2 / 6 (1 N/A)
### 1 · Mechanistic literacy — clear failure
> "in the same cancer category as cigarettes and asbestos" · "A ham sandwich is as dangerous as smoking."
Why: An IARC group states how strong the evidence is that something
can cause cancer. It does not state how much cancer it causes. IARC
(2015): "this does NOT mean that they are all equally dangerous."
### 2 · Quantitative and comparative literacy — clear failure
> "raises your risk of bowel cancer by 18 percent" · "nearly one in five people who eat the stuff will get cancer from it"
Why: 18% is a relative increase, not one in five. On the U.S.
lifetime baseline of about 3.9% (NCI SEER, 2021–2023 data),
3.9% × 1.18 ≈ 4.6%.
The finding underneath survives: processed meat does cause colorectal cancer (IARC, 2015), and the risk rises with the amount eaten. Full report.
A claim about crime statistics
A common sentiment about crime.
# ClaimCheck — "Crime has never been this bad"
**Summary:** clear failures: 3 · partial: 3 · passed: 0 · N/A: 1
**Score:** 1.5 / 6 (1 N/A)
### 2 · Quantitative and comparative literacy — clear failure
> "More than a million violent crimes were reported last year." · "It is the worst it has ever been in this country"
Why: A count with no rate and no earlier year. FBI: 1,119,768
violent crimes in 2025, 327.6 per 100,000 people. In 1991:
1,911,767, and 758.2 per 100,000. The 2025 rate is about 43% of the
1991 rate.
### 3 · Scale and boundary literacy — partial
> "In my city, murders went from 30 to 40 in a single year." · "what is happening here is happening everywhere."
Why: The local rise is a real concern. The national figure moved
the other way: murders fell an estimated 18.1% from 2024 to 2025
(FBI).
The local question survives: the city’s own counts, as a rate, over ten years. So does the point about unreported crime: about half of violent victimizations were not reported to police (BJS, 2024). Full report.
A claim that mostly holds (illustrative)
A short post advising LED swaps, citing DOE.
# ClaimCheck — swap the last incandescent bulbs
**Summary:** clear failures: 0 · partial: 2 · passed: 5 · N/A: 0
**Score:** 6 / 7
### 6 · Recency and updating literacy — partial
> "lighting is about 15% of an average home's electricity"
Why: The figure is from a 2010 DOE guide. EIA's 2020 survey puts
lighting near 6% of home electricity. The advice survives.
No invented faults. Five dimensions pass and say so. Full report.
When to use which
| You have | Use |
|---|---|
| A claim, post, or argument: does it hold up? | ClaimCheck |
| Prose that reads padded or off-register | Smell Check |
| A plan or decision still on the table | TemperPass |
| A finished README, skill, or site: can a newcomer use it? | Cold-eye |
| Files after edits: what went stale? | Detangler |
A loud post with sound numbers passes ClaimCheck and still deserves a Smell Check. A calm post with no denominator fails ClaimCheck and passes Smell Check.
ClaimCheck is a literacy review. It does not rule a claim true or false, and it does not rewrite it.
Built by Catalyst Forge LLC. MIT.