Claude for Journalists: Source-Led Research and Editorial Review

Claude can speed up a reporter's research, help structure an interview and turn rough notes into a reviewable first draft. The catch is that it does not have access to your sources, cannot verify quotes and will sometimes fill gaps with plausible-sounding errors that look like facts.

Used carefully, Claude reduces the mechanical parts of the job — pulling themes from a pile of documents, finding gaps in a question list, making a rough draft readable — so you can spend more time on work only a reporter can do. Tools like fenbs keep verified references and reporting tasks outside any one chat session, so they travel with the story rather than disappearing when the conversation ends.

What Claude brings to a reporting workflow

Claude is strong at reading and synthesizing long text. Give it a 300-page regulatory filing and it will summarize the key sections and flag inconsistencies worth checking. Give it a background briefing and it will suggest questions a subject might find uncomfortable to answer.

What it cannot do: access live sources reliably, speak to confidential contacts, verify that a figure or name is current, or remember anything between separate conversations. Each session starts fresh, which matters when a story stretches across weeks.

Like Claude Code, the standard Claude assistant works best when given clear context and a specific task. Vague prompts produce vague results. The sharper your question and the clearer the material you paste, the more useful the output.

Preparing for an interview

Start by pasting in background material you have already verified: previous coverage, public statements, regulatory filings, annual reports. Ask Claude to identify the key claims the subject has made and the places where those claims conflict or need updating.

Then ask it to draft a question list ordered from broad to specific. Review every question before the interview. Claude may produce a question based on a date or figure it inferred rather than found in your source material — check those against the original documents before you walk in.

The checked question list is worth saving. If you hand the story to a colleague or return to it after a break, a list with source references is a better starting point than another reading of the background pile.

Document analysis

Journalists routinely deal with documents that take hours to read in full: court transcripts, FOIA releases, financial disclosures. Claude can work through them quickly and produce a map of which pages contain the substance you need.

Ask it to extract every named figure, every dollar amount and every date, and to flag where similar information appears with different numbers. That is not verification — you still need to read the pages it flags — but it can narrow a 400-page document to thirty pages worth careful attention.

Claude does not retrieve content from URLs in offline mode. If you paste a document, work from what you pasted. If you need it to reference a web page, paste the relevant text yourself and note the source clearly so you can find it again.

A useful habit is to label every pasted extract with its source and date: the document name, the section and the year it was filed or published. When you return to the story days later, or hand it to a colleague, a labeled extract makes the research navigable. An unlabeled paste might be from a 2021 quarterly report or a current press release — and those can produce very different answers to the same question.

Claude also handles interview transcripts well. Paste a verbatim transcript and ask it to identify the three most specific claims the subject made, or the places where their answers contradict an earlier statement. That is not quoting the subject — it is building a structured reading list from a long document, the same work you would do by hand with a highlighter.

Drafting and editorial review

Claude is useful for first drafts of standard story shapes: news briefs, explainers, profile introductions. Describe the story, paste your checked notes and ask for a draft. Expect to rewrite rather than polish — the draft will reflect the notes you gave it, not the judgment a reporter applies to what matters most.

It is also useful for structural review. Paste a draft and ask which paragraphs are doing the least work, where the sourcing is thinnest, or whether the opening buries the news. Those questions produce useful feedback precisely because the model knows the story only from what you gave it.

Everything Claude produces must go through your normal editorial process. A generated draft does not have sources; you do. Every claim that needs a source still needs one, whether the sentence came from Claude or from your own notes.

Keeping verified references between sessions

The biggest practical problem with using Claude for ongoing stories is that nothing persists between sessions. A week later, the conversation is gone. If you had a question list with source references, a list of checked figures, or a note about a fact you confirmed, you have to reconstruct it.

A story task in fenbs solves this at the project level. Record the verified references as links, note which figures you confirmed and where, and save the question list alongside the source that informed each question. The next time you open Claude for that story, paste the relevant task context so the session starts where the last one ended.

This is different from storing source material in Claude or relying on Claude's memory features. Those are chat-level tools for a single session. A task in a project record travels with the story and is readable by any assistant you use, not just Claude.

Recording correction lessons

Every story where Claude produced a significant error is worth a short note: what it got wrong, why that matters for your beat, and what to check next time. These are calibrations for how you use it on similar stories, not marks against the tool.

Say Claude summarized a financial filing and missed a key liability because it was buried in a footnote. The lesson is: always ask it to check footnotes specifically when reviewing financial disclosures. Write that down somewhere you will find it before the next financial story.

Over time, a short list of beat-specific lessons shapes a much better starting prompt for each session. You spend less time re-explaining what the model tends to miss and more time on the actual story.

What to keep out of Claude

Never paste information that identifies a confidential source: names, contact details, descriptions that could identify them, or the substance of what they told you. Claude's conversation data is handled by Anthropic; source protection is a professional obligation that no AI tool currently covers.

Embargoed material, off-the-record statements, and unpublished court filings belong in the same category. When in doubt, describe the situation in general terms and work from your own checked notes rather than pasting documents whose distribution is restricted.

fenbs does not import or store your Claude conversations. Notes you add to a story task are ones you write yourself. Source contact details belong nowhere in a project management tool.

How to do this with fenbs

The setup is a short loop: save, paste, work, update. Once you have done it for one story, it becomes the normal way to open any Claude session for ongoing work.

1. Create a free account. Add your current story as a task, write what it is trying to establish, and note where it stands right now.
2. Save the primary sources you have already confirmed as links in the task. If you have verified a figure or date, note it in the task along with where you checked it.
3. Write one or two standing rules in Decisions and rules: for example, "always check financial figures against the primary filing." These appear first when a connected assistant reads your project context.
4. When you open a Claude session for the story, paste the relevant task context — what is settled, what still needs checking, where the story stands — so you are not starting from scratch.
5. Before you close the session, update the task with what you confirmed, what you checked and what remains open. The next session, whether in Claude or another assistant, starts from there.

If you use Claude Code or another MCP-compatible tool on reporting projects, you can connect it to fenbs from the integrations page so it reads your project context automatically at the start of each session.