CCAO-F 2.3 — Fact-checking and validating Claude’s work

The fact-checking loop

List the claimsevery figure, date, name, quote, reference
Rank by stakeswhich errors would actually hurt?
Go to the sourceprimary document or system of record
Compare exactlysame number, date, scope, wording?
Correct and logfix it, note what you checked

Claude can help with the first two steps — listing claims and their sources. The third and fourth steps are where independence comes in, and they are yours.

Step one: make the claims visible

You cannot check what you have not isolated. Before verifying anything, turn the prose into a list of claims. Claude is good at this part — it is a reorganisation task, not a truth judgment — and asking for the source of each claim at the same time shows you at a glance which ones have nothing behind them.

A claims table to check againsttext

List every factual claim in the draft above as a table
with these columns:

1. Claim (as stated)
2. Type: figure, date, name, quote, legal or policy
   reference, or other
3. Where it came from: document name and page, web
   citation, calculation, or "general knowledge"
4. Exact supporting text, quoted, if it came from a
   document

Do not add new claims or correct anything yet.

Then rank. A slide of brainstormed ideas needs little checking; a figure going into a board paper needs its own source. The stakes question — which of these would cause real harm if wrong? — decides where your time goes. Deciding when a person must formally review the output is covered in 2.4.

Choose the right check for each claim

Different claims need different checks. The mistake is using one technique for everything — usually re-asking Claude — when each kind of claim has a natural source of truth.

Claim type	Where the truth lives	How to check
Figure from your document	The document itself	Find the quoted passage; confirm number, unit and period
Calculation	The raw data	Recompute a sample; reconcile totals to the source
Internal business fact	System of record (CRM, HR, finance)	Look it up there, not in a slide or an email
Web-cited statistic	The original publisher	Open the citation; follow it back to the primary study
Law, regulation, policy	The official text	Regulator or official publication; check it is current
Quote	Original transcript or recording	Match the words and the context

How much a check is worth

Stronger evidence at the top

Primary sourceofficial text, audited report, system of record
Reputable secondarycites the primary; dated and attributable
Claude’s citationshows where it looked; open and compare
Claude re-checking itselfcatches slips; not independent
Claude sounding confidentnot evidence at all

Checks near the top are independent of Claude’s output. The bottom two stay inside it — useful for catching slips, never enough on their own for a claim that matters.

Web citations deserve a specific habit. With web search on, Claude’s answers include citations and links, and Research runs many searches and returns a report with citations designed to be easy to check. The help centre advises cross-referencing cited sources and using authoritative ones for critical decisions. In practice that means opening the page, finding the sentence, and checking its date and scope — and if the page is itself a summary of someone else’s work, following it back to the original.

Worked example · The regulatory deadline that was only proposed

Omar, a compliance analyst at a regional bank, uses Research to prepare a briefing on a new reporting rule. The report is thorough, cited, and says firms must comply “by 1 March next year”.

He lists the high-stakes claims: the compliance date, which firms are in scope, and two specific reporting thresholds.
He opens the citation for the date. It is a real law-firm blog post — but it summarises the regulator’s consultation paper, not the final rule.
He follows it back to the primary source on the regulator’s own website. The final rule was published later, with a different compliance date and one threshold changed.
He checks the scope claim the same way. It matches the final rule.
He corrects the briefing, cites the final rule directly for every date and threshold, and records which document and version he checked against.

Takeaway: The citation was real and the page said what Claude reported. The problem was one link further back: a secondary source describing an earlier draft. Only going to the primary source caught it.

Validating numbers and analysis

When Claude works with data, much of the checking is reconciliation. With the code execution and file creation capability turned on, Claude runs real code in a sandbox to analyse uploaded data and build spreadsheets with working formulas. That makes arithmetic slips less likely, but it does not make the analysis right: the code can faithfully calculate the wrong thing — the wrong date range, gross instead of net, duplicates left in. Validation asks whether the inputs, definitions and totals match reality.

Worked example · Store growth figures that didn’t tie out

Lena, a finance manager at a retail chain, uploads a sales export and asks Claude to calculate year-on-year growth for each of 48 stores, with a summary for the regional directors.

Row count. She asks Claude how many rows and stores it read. It reports 48 stores — matching her list.
Totals. The total sales figure is 3% higher than the finance system shows for the same period. Something is included that shouldn’t be.
Definitions. She asks how returns were treated. Claude’s analysis used gross sales; the finance system reports net of returns. The code was correct; the definition was wrong.
Spot-check. After a rerun on net sales, she recomputes three stores by hand — the largest, the smallest and one newly opened — and all three match.
She notes the definition (“net sales, returns deducted”) at the top of the summary so readers know what the figures mean.

Takeaway: Every calculation was arithmetically right and the first answer was still wrong. Reconciling to the system of record is what exposed it.

Validating Lena’s analysis

Passes: All rows and stores read48 of 48
Fails: Totals tie to the finance system3% high on first run
Fails: Definitions match the business’sgross used; finance reports net
Passes: Hand-checked sample matcheslargest, smallest, newest store
Missing: Definitions stated in the outputadded before sending

Most data errors are caught by reconciling to something you already trust, not by re-reading the analysis.

Prompting techniques that make checking faster

Anthropic’s guidance on reducing hallucinations includes several techniques that double as validation aids, because they make the output easier to check. They reduce errors; they do not remove the need to check.

Quotes first. For long documents, ask Claude to extract the word-for-word passages relevant to the question before it answers. Anthropic’s long-context guidance says this helps Claude focus on the relevant parts — and it gives you the exact text to verify.
Verify with citations. Ask Claude to find a supporting quote for each claim in its draft, and to remove any claim it cannot support.
Restrict the sources. Tell Claude to use only the documents provided, not general knowledge, so every claim has somewhere to trace to.
Permission to say “I don’t know”. An explicit instruction to flag missing information turns a guess into a visible gap.
Compare runs. Asking the same question more than once and comparing can expose claims that change between runs — a warning sign.

Asking for a self-audit that you can check

Asks for reassurancetext

Is everything in your summary accurate?

Asks for evidencetext

For each claim in your summary, find a
direct quote from the attached documents
that supports it, and give the document
name and page.

If you cannot find a supporting quote,
remove the claim and mark where it was
with [ ].

Finally, list every claim you removed.

The left prompt asks for reassurance. The right one produces evidence you can check in minutes, and a list of what was removed.

Traps the wrong answers are built from

Tempting but wrong	Do this instead
Asking Claude to confirm its own answer	Check against a source outside the output.
Treating a citation link that opens as verified	Find the sentence, check its date and scope, and follow it back to the original.
Finding a second article that agrees	Trace both to the primary source; they may repeat the same error.
Trusting analysis because code ran it	Reconcile totals and definitions to the system of record and spot-check by hand.
Checking every claim with the same effort	List the claims, rank them by stakes, and verify the high-stakes ones first.

You should now be able to

Turn a response into a list of checkable claims with their stated sources.
Match each claim type to its source of truth and the right check.
Rank evidence by independence, from primary source down to Claude’s confidence.
Validate data analysis by reconciling totals, confirming definitions and spot-checking.
Use grounding prompts — quotes first, verify with citations, restricted sources — to make checking faster.

Fact-checking and validating Claude’s work