Best AI Prompts for Long-Form Journalism (2026)

Prompts for long-form journalism#

Act as a features editor. Below are my interview transcripts and documents for a reported article on [DESCRIBE YOUR STORY AND ANGLE]. Using only this material, propose a structure with a nut graf, five or six sections and the strongest scene to open with. For every section, list the transcript or document passage that supports it, quoted exactly, and flag any section where my reporting is thin. [PASTE YOUR TRANSCRIPTS AND DOCUMENTS]

Act as a skeptical fact-checker. Read my draft and list every checkable claim: names, titles, dates, figures, quotes and causal statements. For each, say which of my attached source materials supports it, quoting the passage, or write "no support in the material provided". Do not use outside knowledge and do not guess. [PASTE YOUR DRAFT AND SOURCE NOTES]

Act as an investigative editor. From my notes on [DESCRIBE YOUR SUBJECT], build a timeline in a table with date, event, source document and confidence, and list the conflicts between sources. Mark any date that appears in only one source, and list the questions I should still put to each person named.

Act as a standards editor. Check my draft for attribution problems: unnamed sources without a stated reason, claims stated as fact that rest on one person's account, quotes that may be paraphrased too closely, and places where the reader cannot tell who said what. Suggest rewrites that name the source or state the limit of what is known. [PASTE YOUR DRAFT]

Act as a line editor for long-form features. Tighten the draft below without changing any fact, quote or attribution. Cut repetition, fix pacing, and mark every sentence you changed so I can review it. Keep my voice and the first-person passages. If a change might alter meaning, leave the original and add a comment instead. [PASTE YOUR DRAFT]

What to add to any journalism prompt#

• Your own source material, pasted in. Gemini's guide says to include the instructions and information the model needs to solve the problem, and OpenAI says added context can ground answers in data outside the model's training.
• Delimiters between documents, such as XML tags around each transcript, so the model can tell sources apart. OpenAI's guide documents XML tags for marking where reference content begins and ends.
• An instruction to quote first. Anthropic advises asking for relevant quotes from long documents before the task, which keeps the answer tied to the text.
• Constraints on what it must not do, such as no outside facts, no invented quotes and no changes to attributions. Gemini's guide lists constraints as a core part of a clear prompt.
• The output format you want, such as a table of claim, source and quote. Gemini's guide says you can specify the format of the response.
• Your newsroom's policy on AI use, if you have one. Many policies stress human supervision and verification, according to a 2026 CNTI briefing.

A worked example#

Act as a skeptical fact-checker. Read my draft and list every checkable claim: names, titles, dates, figures, quotes and causal statements. For each, say which of my attached source materials supports it, quoting the passage, or write "no support in the material provided". Do not use outside knowledge and do not guess.

<draft>
[Full text of a 3,000-word feature on a county water-board contract dispute]
</draft>

<source id="1">Transcript of 4 September interview with the board chair, 52 minutes</source>
<source id="2">Contract PDF, text extracted, 38 pages</source>
<source id="3">Board meeting minutes, 12 June and 10 July</source>
<source id="4">My notes from a phone call with the former contractor</source>

A good result for this task:

• Lists claims as separate rows, with each quote, date and dollar figure checked on its own.
• Quotes the supporting passage from a numbered source, so you can find it in seconds.
• Says "no support in the material provided" for anything it cannot match, instead of filling the gap.
• Flags claims backed only by one source, such as your phone notes, as single-sourced.
• Leaves the draft unedited. You then confirm each row against the original recording or document yourself.

Do not use a model to supply facts you have not reported, to identify or characterise real people from memory, to write quotes or to paraphrase a source you have not read. Follow your newsroom's AI policy, or write down your own, and say in the piece or to your editor where AI materially shaped the text if your outlet requires it. We could not read AP, Reuters or SPJ guidance directly on 2026-10-22 (see Limits), so check those organisations' current published standards yourself.

The full prompt library is on our prompt library page.

Limits of this guide#

The vendor guides explain how to ground a model in text you provide; none of them says a grounded answer is correct, and Anthropic's guide gives its quote-first technique as a way to help the model focus on relevant content. We tried to read the Associated Press, Reuters, SPJ and Poynter pages, but their sites blocked or did not return the pages to our fetcher, so we cite a research briefing that summarises newsroom policies instead. We did not test these prompts on any model.