Skip to content

Checks

Claim check

Finds every figure, comparison and quote in a draft and checks each against your approved claims or a source you can open, before a buyer does.

Other tools

Add this skill to Claude

Claude takes skills as a ZIP file. Your download has started; if it didn't, use the button below.

  1. 1In Claude, open Customize, then Skills.
  2. 2Choose +, then Upload a skill, and pick claim-check.zip.
  3. 3Ask for it by name, or let Claude use it when a task matches.

Skills need code execution turned on in Claude's settings. On a Team or Enterprise plan, an admin can share it with everyone.

Download the folder

Try it now, without installing

See the skill work on a made-up input before you add it. Nothing is installed, and you can swap in your own input.

Open in the Claude app

Opens a new chat in the Claude desktop app with the prompt filled in. Nothing is sent until you press enter.

Includes the whole SKILL.md and the sample input. Paste it into Claude, ChatGPT or any chat.

Influenced by
Search Quality Rater Guidelines
Needs
The draft
Works with
Claude, ChatGPT, Codex, Cursor and any agent that reads SKILL.md
Version
1.0.0 · Last checked · License MIT
Maintained by
Sabih AhmedContent
SkillStart here

Claim check

Example output

Claims in a landing page section
ClaimStatusCall
Live in 14 daysVerifiedKeep
Cut reporting time 40%UnsourcedCut or source
The leading toolVagueReplace

Finds every figure, comparison and quote in a draft and checks each against your approved claims or a source you can open, before a buyer does.

The skill

The whole skill, in plain text.

Nothing is hidden. Read it, change it to fit your team, or take the whole folder.

Scripts
Optional Python scripts for the counting. It works without them and says which way it ran.
Network
Opens web pages or your tools if your agent can. Works from pasted data if not.

3 test cases

claim-check/SKILL.md

In the folder

  • claim-check/SKILL.md
  • claim-check/scripts/find_claims.py
---name: claim-checkdescription: >-  Finds every number, comparison, quote and research claim in a draft and checks each against the company's approved  claims in company-profile.md or a source the reader can open, then marks it verified, unsupported, unsourced or  forbidden and returns PASS or FAIL with the fix for each. Use it before any page, post, ad or email with a number or  a "best"-style claim goes out, and as the check step inside skills that write copy. Not for tone or wording  (use voice-check).license: MITcompatibility: >-  Works in Claude, ChatGPT, Codex, Cursor and any agent that reads SKILL.md. Opens sources if the agent can browse;  otherwise asks the reader to paste the source text. Python 3 is optional, for finding claims.metadata:  author: SHAKEUP  version: "1.0.0"  stage: check  plugin: start-here  homepage: https://www.shakeuphq.com/skills/claim-check--- # Claim check Makes sure every number and claim in a draft is backed by something a skeptical buyer could open. ## Purpose - One job: list every checkable claim in one draft and give each a status and a fix.- Stage: check. It ends in PASS or FAIL.- It never makes a claim true by finding a source that sort of agrees. A source must state the claim, for the same thing, in the same period. ## Use when / not | Situation | Use || --- | --- || A draft has numbers, comparisons, quotes or "research shows" | this skill || Another skill reaches its check step on copy | this skill, after `voice-check` || The wording is off but the facts are fine | `voice-check` || Adding new claims to the company's approved list | `company-profile`, mode `update` | ## Inputs | Input | Required | Ways to provide, in order | If missing || --- | --- | --- | --- || The draft | yes | a file the agent can read; pasted text | ask for it || Approved claims | no | `## Claims and proof` in `company-profile.md` | every first-party claim is `unsourced` until the reader gives a source || Sources for external claims | no | links in the draft (opened in this run); source text the reader pastes | the claim is `unverified`; ask for the source | ## Profile read `## Claims and proof` (each approved claim, its source, checked and expiry dates), `### Never say`, `## Competitors` (mention policy), and `## Learnings` entries for `claim-check`. No profile: treat every first-party number as `unsourced` and ask the reader for its source, in one message listing them all. ## Steps 1. **Apply learnings** for this skill. 2. **Find the claims.** Run `scripts/find_claims.py --draft draft.md`. Without Python, read every sentence and pick out: numbers with %, x, times or a currency; comparisons (more than, faster than, the best, #1, industry-leading, leading); quoted passages; "according to", "studies show", "research shows"; and any year next to one of those. Also catch claims the pattern misses: "trusted by thousands", "used by top teams", "proven". List every one with its line. 3. **Sort each claim.**    | Type | Means | Needs |   | --- | --- | --- |   | First-party | about the company, its product or customers | a match in `## Claims and proof` |   | External | a fact about the market or the world | a source opened in this run, or text the reader pastes |   | Competitor | about a named competitor | a public source, and the profile's mention policy allows it |   | Customer quote | words from a customer | a match in `## Claims and proof` or the reader's confirmation it is cleared for use |   | Opinion | clearly framed as the company's view ("we think") | nothing; not a claim | 4. **Check first-party claims** against the profile. A match must be the same number, about the same thing, for the same period. Then check the dates: past `Expires` is `expired`; `Checked` over 12 months ago is a warning. 5. **Check external claims.** Open the source. Find the sentence that states the claim. Run `scripts/find_claims.py --claim "<claim>" --source <source text file>` as a first filter: under 30% of the claim's words in the source means it's very likely unsupported. Passing the filter isn't proof: read the sentence and confirm the number, the subject, the period and the population match. Record the excerpt.    | Source problem | Status |   | --- | --- |   | The page doesn't state the claim | `unsupported` |   | It states it for a different group, market or year | `unsupported`, and say what it does state |   | The source cites another source | follow it once; cite the original |   | It can't be opened (paywall, 403, dead link) | `unverified`; ask the reader for the text |   | It's the vendor's own marketing about itself | allowed, but say "according to <vendor>" | 6. **Check what may never be said.** Any match to `### Never say`, or a competitor claim the mention policy doesn't allow, is `forbidden`. 7. **Write the fix** for each claim that isn't verified:    | Status | Fix |   | --- | --- |   | `unsupported` | Replace with what the source does say, or cut the number |   | `unsourced` | Ask for the source; until then, cut the number or turn it into the mechanism ("cuts the steps from 6 to 2" if that is a fact about the product) |   | `unverified` | Ask for the source text; keep the claim out until it's checked |   | `expired` | Ask for a current figure, or date the claim ("in 2024, ...") if the source still stands |   | `forbidden` | Cut it |   | vague superlative ("the best", "leading") | Replace with the specific proof behind it, or cut | ## Check and verdict Severity: `critical` (unsupported, forbidden, unsourced first-party number, expired), `warning` (unverified, a vague superlative, a source checked over 12 months ago), `suggestion` (a verified claim whose sentence could cite its source more clearly). | Verdict | Condition (first match wins) || --- | --- || `BLOCKED` | No draft || `FAIL` | Any critical finding || `PASS WITH NOTES` | Warnings or suggestions only || `PASS` | Every claim verified or opinion | No named stops. The calling skill fixes and reruns. ## Learn - A source the reader confirms for a first-party claim: add the claim to `## Claims and proof` with its source and today's date, and record a `fact` learning.- A source that turned out wrong or dead: record a `fact` learning naming the source, so later runs don't trust it. ## Output ```Verdict: <PASS | PASS WITH NOTES | FAIL | BLOCKED> · Coverage: <complete | partial (sources not opened: n)>Claims found: <n> · verified <n> · unsupported <n> · unsourced <n> · unverified <n> · forbidden <n> · expired <n> | Line | Claim | Type | Status | Evidence or problem | Fix |(critical first; for verified claims, the source and a short excerpt)``` Then the handoff block. ## Hand off ```handoffskill: claim-checkversion: 1.0.0verdict: FAILcoverage: partial (2 sources behind a paywall)for: <draft name>outputs:  critical: 1  verified: 6  needs_source: ["Teams cut CAC by 23%"]next: []open: [source for "Teams cut CAC by 23%"]``` ## Failure modes | Problem | Do this || --- | --- || The draft cites "internal data" | Treat as first-party: it must be in `## Claims and proof` || A number is rounded differently from the source (24% vs 23.7%) | Verified if rounding is normal; flag a rounding up of a range ("up to 40%" from "20 to 40%") as unsupported || An AI answer engine is the "source" | Not a source. Find the page it cited || The same claim appears several times | Check it once; list every line || A testimonial is paraphrased | It's no longer a quote: remove the quotation marks or use the cleared original | ## Influenced by - SHAKEUP's content engine, where no number reaches an article unless its source was fetched and checked in the same run.- Google's Search Quality Rater Guidelines on trust, which treat unsupported and misleading claims as a sign of low quality: https://guidelines.raterhub.com/searchqualityevaluatorguidelines.pdf

How it thinks

Inputs in, a verdict out, and a stop only where you decide.

What we built

In our own content engine, no figure reaches an article unless its source was fetched and checked in the same run. This skill applies the same rule to any draft, against your own approved claims.

It reads your company profile if you have one, or asks for only what it needs. Then it works through its steps, checks its own work and ends in a verdict.

The steps

  1. 01

    Find the claims

    Its script lists every sentence with a figure, a comparison, a quote or "research shows". It also catches softer ones like "trusted by thousands".

  2. 02

    Sort each one

    About your company, about the market, about a competitor, a customer quote, or plainly an opinion. Each type needs a different kind of proof.

  3. 03

    Check against the source

    The source has to state the claim for the same thing, the same group and the same period. If it only sort of agrees, it doesn't count.

  4. 04

    Check what you may never say

    Anything on your "never say" list, or a competitor claim your mention policy doesn't allow, is cut.

  5. 05

    Write the fix

    Each claim that isn't verified gets one fix, from using what the source does say to cutting the figure.

What you need

  • Required

    The draft

    A file your agent can read → Text you paste

  • Optional

    Your approved claims

    The Claims and proof section of your company profile

  • Optional

    Sources for outside claims

    Links in the draft, opened during the run → Source text you paste

What it can return

  • PASS

    Every claim is verified, or is framed as an opinion.

  • PASS WITH NOTES

    Warnings only, such as a source checked over a year ago.

  • FAIL

    Any claim is unsupported, unsourced, expired or on your never-say list.

  • BLOCKED

    There's no draft.

Where it stops for you

It runs through without stopping. You review the result.

Example output

What you get back.

A run for a made-up company. Your output depends on your inputs.

Here's a run on a landing page section for a made-up B2B SaaS company.

Verdict

FAIL

Coverage: complete

LineClaimTypeStatusEvidence or problemFix
1Live in 14 days on averageCompanyVerifiedApproved claim, onboarding data, checked this monthKeep
2Cut reporting time by 40%CompanyUnsourcedNot in the approved claimsAsk for the source, or cut the figure
3140 B2B SaaS companiesCompanyVerifiedApproved claim, CRM count, checked this monthKeep
The handoff block the next skill reads
skill: claim-check
version: 1.0.0
verdict: FAIL
coverage: complete
for: landing page section
outputs:
  critical: 1
  verified: 2
  needs_source: ["Cut reporting time by 40%"]
next: []
open: [source for the reporting time claim]

What good looks like

  • Every figure in the draft appears in the table
  • Verified claims show the source and the sentence that states them
  • Nothing is kept because a source "sort of agrees"
  • Each problem has one clear fix

Install

Use it in the tool you already have.

Pick your tool. Each tab has the shortest path we know of.

  1. 01
    Download claim-check.zip.
  2. 02
    In Claude, open Customize, then Skills.
  3. 03
    Choose +, then Upload a skill, and pick the ZIP. Code execution must be on in your settings.
  4. 04
    Ask for it by name, or let Claude use it when a task matches.

Questions

Before you add it.

What counts as a claim?

Any figure, comparison, quote or appeal to research, plus soft claims like "trusted by thousands" or "the leading". Opinions framed as yours, such as "we think", don't count.

Can it open the sources itself?

Yes, if your agent can browse. Otherwise it asks you to paste the source text, and marks the claim unverified until you do.

Why is a vendor's own page allowed as a source?

It's allowed only for what the vendor says about itself, and the draft has to say "according to" that vendor. It never stands in for an independent source.