Editing and Fact-Checking AI Drafts With Claude

#01Why separate editing into passes?

Because each kind of edit needs a different kind of attention. If you ask Claude to "improve this", it will change a bit of everything: some structure, some wording, some tone. You'll get a smoother draft, but you won't know what changed or why, and real problems can survive.

Single-purpose passes are easier to review. You can accept a structural change without accepting new wording, or reject voice changes while keeping clarity fixes. Anthropic's prompting guide makes a related point: Claude follows instructions precisely, so if you want changes rather than suggestions, say so, and be explicit about what kind of change you want (Anthropic).

Our view: four short passes beat one long one. They also teach you more about your own writing, because you see each type of weakness separately.

#02Pass 1: Does the structure work?

Start with the big picture. Ask Claude to review the draft as a whole, not line by line.

<draft>[paste]</draft>
<brief>[paste or summarise: reader, main question, goal]</brief>

Review the structure only. Don't rewrite sentences.
1. Does the opening answer the main question within 3 sentences?
2. List each section's main point in one line. Are any repeated,
   missing or out of order?
3. Which section is weakest for this reader, and why?
4. Suggest a revised outline if needed.

Fix structure before anything else. Polishing sentences in a section you're about to cut is wasted effort.

#03Pass 2: Is it clear?

Next, ask for line edits focused on clarity: shorter sentences, plainer words, fewer abstractions. This is where Claude is most useful, and where it's easiest to review because the changes are small.

Edit for clarity only. Keep the meaning and our voice.
- Split sentences over 25 words.
- Replace vague words (solution, leverage, robust) with specific ones.
- Turn passive sentences active where it helps.
- Cut filler phrases.
Return the edited text, then a table of the 10 most significant
changes: original, edited, reason.

The table makes the edit reviewable. If you disagree with a type of change, tell Claude and it will stop making it.

#04Pass 3: Does it sound like you?

If you've built a brand-voice project (lesson 4), run this pass inside it. Ask Claude to compare the draft with your voice guide and examples, and to point out lines that don't fit, rather than rewriting everything.

Compare this draft with our voice guide and example pieces.
List up to 10 lines that sound least like us, with a suggested
rewrite for each and the voice rule it breaks.

This keeps your own writing intact and only touches the lines that need it.

#05Pass 4: Are the facts right?

This is the pass that matters most, and it needs a different approach. Ask Claude to extract claims first, then check them, rather than asking "is this accurate?", which invites a confident yes.

Fact-check this draft.
Step 1: List every factual claim: numbers, dates, names, quotes,
product details, comparisons and "studies show" statements.
Number them.
Step 2: For each claim, search for the best source. Open it and
quote the exact sentence that supports or contradicts the claim.
Step 3: Mark each claim Supported, Partly supported, Contradicted
or Not found. For anything not fully supported, suggest a fix:
a better source, a softer wording, or removal.
Don't mark a claim Supported unless you've quoted the source.

The rule that Claude must quote the source is the key. Anthropic's guidance on reducing hallucinations recommends asking Claude to find supporting quotes and to retract claims it can't support (Anthropic). It also recommends explicitly allowing Claude to say it doesn't know, which makes "Not found" an acceptable answer rather than a failure.

With web search on, Claude cites the pages it used (Anthropic). Click through for every claim that matters, especially numbers. A quote can be real but taken from an old version of a page, or from a site that misreported the original.

#06What does Claude get wrong when fact-checking?

Knowing the failure modes helps you check efficiently:

• Secondary sources. Claude may find a blog that quotes a statistic and treat it as the source. Ask for the original study or official page.
• Out-of-date pages. Pricing, product features and regulations change. Ask for the date of each source.
• Near-matches. A source says "up to 40%" and the draft says "40%". Claude may call that supported. Ask it to flag any difference in wording.
• Your own product. Claude doesn't know your product better than your documentation. Keep a fact sheet in your project and tell Claude to check product claims against it.
• Confident summaries. When a source is long, Claude may summarise it more neatly than it deserves. The quote requirement protects against this.

#07How should you handle statistics?

With extra care. A statistic that can't be traced to its original source shouldn't be published, however often it appears online. Ask Claude to trace each number back to the organisation that produced it, and to tell you the sample, date and method if the source gives them.

If you can't find the original, cut the number or replace it with a qualitative statement you can support. Readers trust specific numbers more, but only when they're right, and AI engines increasingly favour content with clear, attributable data. Our guide on statistics, quotes and sources that raise citation rates explains why attribution matters for AI visibility.

#08How do you check quotes from real people?

Treat quotes as facts. Ask Claude to find where each quote was published and to confirm the exact wording and who said it. Misattributed quotes are common online, and a famous name attached to a line it never said can damage credibility quickly. For quotes from your own interviews or customers, check against your recording or notes, and get approval for the final wording.

#09What does a pre-publish checklist look like?

Turn the passes into a checklist you run on every piece. Ask Claude to run it and report back in a table.

Run our pre-publish checklist on this draft:
1. Opening answers the main question in 3 sentences or fewer
2. Every H2 starts with a direct answer
3. Every statistic links to its original source
4. Every product claim matches our fact sheet
5. Every quote is attributed and verified
6. No banned phrases from our voice guide
7. Links work and point to the right pages
8. Meta description under 155 characters
Report pass/fail for each with the line that fails.

Save it as a Skill so it loads whenever you ask Claude to check a draft. Lesson 4 explains how.

#10From MentionX data: why accuracy affects AI citationsOriginal data

Method. Read-only aggregate queries over MentionX scans from 22 August to 29 October 2026 (UTC): 2,217 engine runs (2,154 completed) across 18 anonymised projects and six engines.

Accurate, well-sourced pages are the ones AI engines can safely quote. MentionX's cross-engine aggregates show which kinds of pages get cited most often:

Our view: a fact-check pass is an AI-visibility task as well as a quality task. To check which sources AI answers cite for your topic, see our AI citations checker guide.

Limits. A convenience sample of mostly B2B software projects, run through engine APIs, which can behave differently from the consumer apps. Perplexity, Copilot and DeepSeek figures come from a single day (26 October 2026), so treat them as a snapshot, not a trend.

#11Worked example: Ledgerline checks a payments guide

Ledgerline is a fictional accounts-payable tool. Figures are illustrative.

Ledgerline's editor runs the four passes on a guide to supplier payment terms. The structure pass moves the definition of payment terms to the top. The clarity pass shortens fourteen sentences; she rejects two changes that lost nuance.

The fact-check lists 23 claims. Claude marks 18 Supported with quotes, three Partly supported, one Not found and one Contradicted. The contradicted claim is a late-payment rule that changed last year; the official government page has the current rule. The Not found item is a statistic from a blog that she can't trace, so she cuts it.

The guide ships with every number linked to its original source. Ledgerline is fictional and the details are illustrative.

#12Should you tell readers you used AI?

That's a policy decision for your organisation, and expectations differ by audience and industry. What matters most is accountability: a named person or team should stand behind every published piece and be able to explain where its facts came from. A clear sourcing standard, and the checks in this lesson, do more for trust than a label on its own.

#13What should you try before the next lesson?

Take a recently published article and run Pass 4 on it. Most teams find at least one claim they can't source. Fix it, then save the fact-check prompt and the checklist to your library.

Next: analysing CSVs from GA4 and ad platforms. Or return to the Claude for Marketers track.

#14How do you edit someone else's writing with Claude?

Carefully, and with their voice in mind. When you edit a colleague's or a guest author's piece, the aim is to make it clearer without making it sound like your brand, or like Claude. Tell Claude whose piece it is and what to preserve: "This is a guest post by a finance director. Keep her examples, opinions and phrasing. Only fix clarity, length and factual issues."

Share the change table with the author rather than a silently edited draft. People accept edits far more readily when they can see what changed and why, and they'll often catch a place where an edit changed their meaning. For expert contributors, ask Claude to list any technical statements it softened or reworded, so the author can confirm the new wording is still correct.

#15How do you handle sensitive or regulated topics?

Some content carries more risk: health, finance, legal topics, and claims about safety or performance. For these, Claude's fact-check is a starting point, not a sign-off. Ask Claude to flag every sentence that could be read as advice or a guarantee, and every comparison with a named competitor, then send those to the right reviewer in your organisation.

It also helps to ask Claude to point out where a reader could misunderstand a claim. "Up to 50% faster" might be accurate and still misleading if most customers see far less. A short note on what the figure depends on often fixes the problem without losing the point.

#16How do you keep a record of your checks?

Save the fact-check table with the published piece, in your project or your content system. It records which source supported each claim on the date you checked. When a reader questions a figure, or a source changes, you can find the evidence in seconds. It also makes updates much faster: when you refresh the article next year, ask Claude to re-check only the claims marked as time-sensitive, using the saved table as a starting point.

Over time, the tables also show patterns. If the same type of claim keeps failing, such as old market-size figures or vague "studies show" lines, add a rule to your voice guide so drafts stop including them in the first place. Good editing isn't only about fixing one draft; it's about making the next draft better before anyone has to fix it.

#17What can't Claude check for you?

Some things need a person with context. Claude can't tell whether a customer has approved a quote, whether a product feature has been delayed since the documentation was written, or whether a claim fits what your legal team agreed last month. It also can't judge whether an argument is fair to a competitor or whether an example might embarrass a partner. Build those checks into your sign-off: one named person confirms product claims, and one confirms anything involving customers or competitors. Claude makes their job faster by giving them a clean list of exactly what to look at. That split keeps reviews short and accountability clear.

#18Common questions

Can Claude fact-check an article? Yes, as a first pass. Ask it to list every claim, find and quote a source for each, and flag anything unsupported. Then check the flagged and important claims yourself.

How do I stop Claude rewriting my whole draft? Ask for one type of edit at a time and request a table of significant changes, so you can accept or reject each one.

Does Claude make mistakes when checking facts? It can. Common issues are secondary sources, outdated pages and near-matches in wording, so require quotes and check dates.