Claude can speed up the organisational parts of reporting: planning an interview, listing every factual claim in a draft and what would verify it, comparing quotes against transcripts, and finding a spine for a pile of notes. It cannot report. It has spoken to no one, cannot verify a claim against the world, and will produce a plausible quote or statistic if you let it fill a gap. Treat it as a fast, literal research assistant that works only from material you give it, marks everything else unverified, and never supplies a fact or a quote.
Auf dieser Seite
- What is the one rule that makes this safe?
- How do you prepare an interview with it?
- How do you check facts when it cannot check facts?
- How do you get structure out of a pile of notes?
- How do you audit quotes against your recordings?
- What are the ethical lines?
- Where does Claude fail at this?
- Which journalism skill fits which job?
- How do you install a journalism skill in Claude?
- All-in-one: every skill, one price
- Related guides
- Frequently asked questions
Journalist AI Skill - Peregrine
Get Journalist AI Skill - Peregrine →What is the one rule that makes this safe?
Claude never supplies a fact or a quote. It handles material you provide and marks everything else.
A language model generates likely text. In most writing that is a feature; in reporting it is the hazard, because a likely quote attributed to a real person reads exactly like a true one, and the model has no internal flag separating the two. You supply the flag: input discipline (give it transcripts and documents labelled by source), output discipline (every claim and quotation carries its source tag, and anything without one is UNVERIFIED or removed), and verification (you check a sample of tags against the originals every time, because a tag is only a claim of provenance). Everything below follows from this. For verification as a discipline see our fact-checking guide; for background reading, the research assistant guide.
How do you prepare an interview with it?
Interview prep is where Claude is most useful and least risky, because the output is questions, not claims. The craft rules are old: research first, start with what the subject answers easily, put the hardest question late but not last, ask for specific moments ("walk me through that Tuesday") rather than opinions, and ask who else to speak to.
The claim map separates what you can prove from what you were told, which is easy to lose when tired. Forbidding pre-written answers matters: a model that drafts the subject's replies leaks your expectations into the questions. The pressure-points list is the most valuable output, since it forces you to find the document before the interview rather than after.
How do you check facts when it cannot check facts?
Claude cannot verify anything against the world. It can read your draft, list every checkable statement, say how well each is supported by the material you provided, and name the document that would settle it. That turns "I should fact-check this" into a ledger.
An illustration of the output (invented content, shown for format only):
Row F07 shows the use: the model caught a gap between "a third" and the figures in the source. Do that arithmetic yourself too.
The skill this guide is built on Journalist AI Skill - Peregrine - $19 Builds a feature from what is actually sourced. No quote or claim is written into a draft unless it is directly sourced or marked as an unverified placeholder. Covers nut-graf placement, narrative structure, interview-question design, source attribution and a fact-check table of gaps. It is not a fact-checker with independent access to primary sources. Get Peregrine →
How do you get structure out of a pile of notes?
The common first-draft failure is chronology: everything in the order you learned it. A feature usually needs an opening moment, a nut graf (the paragraph saying why this matters now), and a middle organised by tension rather than date.
Two structures is deliberate: you rarely recognise the best one until you see the alternative. The gaps list is often longer than writers want, and honest. Where it writes [REPORTING NEEDED], resist prompting it to "just fill it in".
How do you audit quotes against your recordings?
Before you file, paste the draft and transcript and have Claude compare every direct quote. This is mechanical, and the model is good at it if told to be literal.
Run this on every piece, your own careful ones included. Fatigue changes quotes.
What are the ethical lines?
Quotes are real or they are not quotes. An AI-written line in a person's mouth is fabrication however reasonable it sounds. Placeholders stay marked in the working file and are replaced with a real quote before filing.
Source protection comes first. Before pasting anything into any AI service, consider what you expose: a confidential source, leaked documents, unpublished findings. Check your organisation's policy and the provider's data terms, and redact identifying details where a source could be harmed.
Disclose according to your outlet's rules, and ask your editor before the piece runs, not after.
Legal review stays human. Claude can flag a damaging claim about a named person resting on one source. It cannot tell you whether the claim is defensible where you publish, and this guide is not legal advice.
The byline carries responsibility. A model being wrong is not a defence.
Investigative Journalist AI Skill - Ottoline
Get Investigative Journalist AI Skill - Ottoline →Where does Claude fail at this?
It fills gaps with plausible text. Ask for "a quote that captures her view" and it will write one; ask for a statistic and you get a believable number. Prevent it with labelled sources.
It misattributes. In a multi-speaker transcript it can assign a line to the wrong person, especially after interruptions. The quote audit catches this.
It "verifies" from memory. Asked "is this true?", it answers from training data that may be out of date or wrong, in a confident tone. Familiarity to a model is not evidence; only documents and people count.
It flattens hedges. "He may have been aware" becomes "he was aware" in a summary. Check modal verbs ("may", "reportedly", "according to") against the source, because defamation risk enters there.
It cannot weigh a source. It does not know who benefits from a claim, and it carries no memory of what is off the record unless you label every item.
It writes in an average news voice. A distinctive feature needs your observation. You were in the room; it was not.
Which journalism skill fits which job?
| Skill | Best for | Price |
|---|---|---|
| Journalist AI Skill - Peregrine | Profiles, features and narrative nonfiction built from sourced material | $19 |
| Investigative Journalist AI Skill - Ottoline | A story that starts as a tip: triage, source protection, verification plan | $19 |
| All-in-one | Both, plus every other skill in the catalog | $12/mo or $69/yr |
How do you install a journalism skill in Claude?
With an All-in-one subscription there is nothing to download. Open Customize → Connectors, click +, choose Add custom connector and paste https://mcp.kissmyskills.com/mcp. Sign in with your email, then ask for a skill by name, for example "load Peregrine". The same connector works in ChatGPT, Claude Code, Codex and Cursor; setup steps are on the connector page. Signing in is free, and your first skill call starts a 7-day free trial with every skill.
Bought the single file? A .skill file is a ZIP archive with SKILL.md inside. In Claude, turn on code execution under Settings → Capabilities, then Customize → Skills → + → Upload a skill, and toggle it on. In ChatGPT, Gemini or another chat, paste the contents of SKILL.md into custom instructions.
All-in-one: every skill, one price
All-in-one is $12 a month or $69 a year, with 7 days free and every skill included, and a 30-day money-back guarantee. Read why a subscription, see the All-in-one plan, or open the connector setup.
Related guides
- How to Use Claude for Fact-Checking
- How to Use Claude as a Research Assistant
- How to Write a Press Release With AI
- How to Use Claude as a Travel Writer
Frequently asked questions
Is there a Claude skill for journalists?
Yes. Journalist AI Skill - Peregrine ($19) builds features from sourced material, writes no quote or claim unless it is sourced or marked as an unverified placeholder, and keeps a fact-check gap table. Investigative Journalist AI Skill - Ottoline ($19) covers the path from a tip to a manuscript, including source protection and a two-source verification plan. Both are included in the $12/month All-in-one plan.
Can Claude fact-check an article?
Not by itself. It cannot verify claims against the world. It can read your draft, list every checkable statement, compare each with the source material you provide, and flag unsupported claims, single-source claims and quotes that do not match your transcript. Verification still comes from documents and people you contact.
Will Claude make up quotes?
It can, if you let it fill a gap, and the result reads like a real quote. Prevent this by giving it labelled source material, requiring every quotation to be reproduced word for word with its source, and running a separate quote audit against your transcript. A line the person did not say or send you is not a quote.
Can Claude help me prepare for an interview?
Yes. Give it your background material and what you can and cannot document. It can produce a claim map, a question plan that places hard questions late, follow-ups for vague answers, and a list of documents to hold up. It should not write the subject's expected answers or add facts about them from general knowledge.
What is Claude bad at in journalism?
It fills gaps with plausible text, misattributes lines in multi-speaker transcripts, treats familiarity as proof when asked if something is true, softens hedged language into flat claims, and cannot judge source reliability, off-the-record status or legal risk. It also cannot do the reporting: calls, records requests and finding the second source.
Journalist AI Skill - Peregrine
Or get it with every other skill in All-in-one: $12/month, 7 days free, every skill.