AI in the newsroom: a journalist's guide

A journalist's day is made of two different jobs. One is the job you're paid for: finding what nobody's saying, calling the right person, asking the uncomfortable question, and writing it so it's both true and readable. The other is everything around it — reading thirty press releases that contain one piece of news, transcribing forty minutes of an interview, cutting a piece by eight hundred characters, producing three social variants, and tracking down when that company got that subsidy. That second job eats half the day and nobody ever notices it.

AI is good at exactly that second job — and dangerous the moment you let it near the first. This piece draws that line properly and then shows what's underneath it: how to have background research done in a way that can still be verified, how to read public data on contracts and subsidies, how to trim an intro to an exact character count, when to draw a diagram and when never to generate an image, and what must never end up in an outside system, even if the deadline is twenty minutes out.

This is the hub for the whole journalism section. Every part has copy-paste prompts — just fill in the brackets — and links to more detailed guides that go deeper on individual steps. You don't have to read it in one sitting; start with the section on protecting sources, then come back to whichever part of the day you're in.

A typical scenario

Klára covers news for a regional newsroom. At half past eight in the morning, her inbox has three press releases: the regional authority is touting a new subsidy for school repairs, the opposition's release calls it an election ploy, and the third is from the company that will do the repair work. She owes two pieces by lunch, has an interview lined up at two with the principal of one of those schools, and the deadline is at five.

She knows exactly where her day goes, because she once tracked it for a week.

Press releases, 40 minutes. Reading three releases about the same thing, telling what's actually new, and finding where they contradict each other. The contradiction is usually the story. Klára does it by hand: reading, underlining, retyping the numbers into a note.

Digging, 50 minutes. How big the subsidy is, where the money comes from, whether the region got a similar one before, who won the contract and for how much. Clicking through registries is slow, and she has to retype the numbers by hand.

Trimming, 25 minutes. The piece runs 4,600 characters; the section only fits 3,800. Cutting your own text is the slowest job in the world, because every sentence looks necessary.

Transcribing the interview, 60 minutes. Forty minutes of recording means an hour of transcription, then another half hour hunting for the one spot where the principal said something quotable.

Social, 20 minutes. Turning the finished piece into a Facebook post, something shorter for X, and a line for the newsletter.

That's nearly three hours a day in which nothing gets found out. Klára needs to cut that in half — not to file more pieces, but to have time to call the other side before it goes out. The rest of this guide walks through her day step by step, and we'll come back to it at the end.

Where AI helps in a newsroom, and where it doesn't

I'm writing this part first and flatly, because everything else follows from it.

The hard line

AI is not a source. A language model doesn't know what happened. It can produce text that sounds like it knows — and it sounds exactly as confident whether it knows the thing or is making it up. The sentence “the ministry says it was 340 million” looks the same in a reply whether the ministry actually said it or it's a number generated as a plausible continuation of the sentence.

That gives you a rule with no exceptions in the newsroom: no fact may enter your text from the model unless you've seen it yourself at a traceable source. Not a paraphrase, not a number, not a date, not a name, not a quote. AI may tell you where to look for the fact. Verifying it is on you.

The other half of that line: a named person signs off on the piece. Not the newsroom, not the tool — a person who stands behind every sentence and can defend it when the other side's lawyer calls. When a mistake makes it into print, the model isn't accountable — you are. That's the one thing that makes what you write journalism instead of content.

So where does it help

Three areas where AI genuinely saves time in a newsroom:

Finding. Where to look, which database tracks this, who covers it. The model is fast orientation in unfamiliar terrain, not an answer — what it hands you are leads you go and walk yourself.

Sorting. Thirty pages of material, three press releases, an hour of transcript, a list of two hundred contracts. Sorting it into “this is new,” “this contradicts that,” and “this is just padding” is mechanical work the model does well and fast. The checking is still yours — but checking sorted material is ten times faster than sorting it yourself.

Form. Headline, intro, length, tone, commas, typography, social variants, captions. Nothing here decides what's true, only how it's packaged — the biggest time savings and the least risk.

Where it doesn't belong

It must not generate facts. If a reply doesn't have a link you can click, it's not information, it's a hypothesis.

It must not write the whole piece. Not on principle, but because of how it plays out: a generated article is about eighty percent true, and the other twenty is spread through it in a way you won't spot. Fixing someone else's text when you don't know what's been verified takes longer than writing it yourself.

It must not talk to a source for you. You can have questions prepared, but you conduct the interview and follow up based on what the person actually says. An automated system wouldn't ask the follow-up that matters, because it wouldn't notice the hesitation in someone's voice.

It must never see a protected source's identity. There's a dedicated section on this at the end, and it's the single most important sentence in this entire piece.

It must not decide what gets published. A model can't weigh who a story would hurt against the public interest. That's an editorial call, and a human makes it.

A rule worth taping above your desk

For every task you hand AI, ask one question: if an intern had done this, what would I double-check? The answer is exactly what you have to double-check after the model too — except an intern will admit when they couldn't find something, and the model will just fill in the gap.

Research and verification

Morning, three press releases, fifty minutes of digging. This is where you can save the most time — and where a whole piece can go off the rails the fastest.

Three detailed guides on this site go deep on this and I won't repeat them here: search with citations breaks down how to ask so the answer rests on sources, and why you have to click through to the citations; deep research shows when it pays to let the model work for twenty minutes on structured research instead of a quick answer; and fact-checking is the manual on what always gets verified, what only sometimes does, and what never goes out without a second source. Read them. This section is just the newsroom cross-section: what to do with all this in the morning.

Turning a press release into facts

A press release is a marketing text pretending to be information. Working with one has three steps: pull out what's actually new, separate that from spin, and list what's missing from it. That third step is the reason you read it at all.

This is a press release. Process it into a structure for a news piece.
Don't infer anything — work only with what's in the release.

1. FACTS: a numbered list of verifiable claims. For each, note who states
   it in the release (institution, named person, "unspecified").
   Copy figures exactly, including units and time periods.
2. SPIN: sentences that look like facts but are actually opinion, a goal,
   or a promise about the future. Quote them verbatim.
3. MISSING: what would logically be in a release on this topic that
   isn't — specifically amounts, deadlines, names of those responsible,
   comparison with the past, the other side.
4. QUESTIONS TO ASK: 5 questions for the spokesperson, ranked by how
   uncomfortable they are. Each must be answerable in one sentence or number.

Text of the release:
[paste the press release]

It returns four blocks, and points 3 and 4 are the valuable ones — the model is decent at noticing that a release about a subsidy doesn't say where the money comes from. Check point 1: models tend to lightly reword numbers and drop the time period attached to them (“340 million” instead of “340 million over three years”). Always read the numbers in the original.

Three releases about one thing

Klára's morning situation: the authority, the opposition, and the contractor are all talking about the same subsidy. The difference between their versions is the story, and it's hard to spot, because reading three texts back to back makes a contradiction in the numbers easy to miss.

You've been given three press releases about the same event from three
different parties. Compare them.

List as a table: CLAIM | release A | release B | release C
For each claim, note what each release says about it — and when one is
silent on it, write "not stated."

Then separately:
- CONTRADICTIONS: where the releases differ on a number, a date, or a
  claim, and what could settle the contradiction (which document,
  which database, who to ask).
- AGREEMENT: what all three agree on — the firmest ground.
- DIFFERENT VOCABULARY: how each side names the same thing.

Don't judge who's right. Just show where they diverge.

Release A ([who]):
[paste]
Release B ([who]):
[paste]
Release C ([who]):
[paste]

The most useful prompt in the whole morning block. One thing to watch: the model sometimes flags a contradiction that's really just different rounding or a different time period. Verify every finding in the originals before you build a paragraph on it.

A briefing on a topic you don't understand

A reporter gets a topic they've never covered and two hours to get up to speed. The goal of this kind of research isn't to write the piece — it's to know enough to tell on the phone when someone's lying to you.

I'm working on a news piece about [topic] in the context of
[region / institution]. I'm not writing it through you — I need to get
oriented in the topic, so I know what to ask and where to verify it.

Put together a briefing:
1. How it works: the mechanics of the thing in 8 sentences, no jargon.
   Where the money is, who decides, what's mandatory and what's voluntary.
2. Who the players are: roles, not names — the contracting authority,
   the contractor, the oversight body, who pays, who it serves.
3. Where the public data is: specific names of registries and documents
   that track this, and for each, what you can learn from it.
4. Common ways mistakes and manipulation happen in this field —
   generally, as warning signs, not accusations.
5. Five questions I need to answer before I write the first sentence.

Use search, and cite a source with a link for every factual claim.
Where you're not sure, say so instead of guessing.

Use this in a tool with search turned on, or you'll get generic knowledge with no way to check it. Point 3 is the reason to do this at all — a list of databases you'd otherwise spend years assembling. And click every link; search with citations calls this the three-click rule, and in a newsroom it applies double.

Verifying a single claim

The most common situation before deadline: your text has one borrowed sentence and you're not sure about it. Verification works by breaking the claim into parts that can each be checked separately.

I want to publish this claim. Before I do, break it down for me.

Claim: "[paste the exact sentence as I want to write it]"
Where I got it: [press release / other outlet / spokesperson / document]

1. Break the claim into verifiable components (number, time period,
   subject, causal link, comparison) and list them separately.
2. For each component, say exactly where it gets verified — the name
   of the registry, document, or annual report. Not "public sources" —
   a specific name.
3. Flag the weakest component — the one most likely to be skewed or
   taken out of context.
4. Write a more cautious version of the same claim, one that still holds
   under narrower conditions, that I can use until I have confirmation.

Don't tell me whether it's true. Show me where to find out.

Point 4 is your safety brake before deadline: you'll often find the cautious version says almost the same thing, and you don't have to take the risk. Check point 2 — when the model names a database you don't recognize, verify it exists before you trust it. Models are especially happy to invent registries that don't exist.

Turning an interview transcript into quotes

A forty-minute recording. A dictation tool or an assistant that processes audio can do the transcript; the remaining work is finding the three spots in six thousand words that will actually go in the piece. Cleaning up the transcript itself is covered in cleaning up a transcript — this is the step after that.

This is a verbatim transcript of an interview I conducted. The speaker is
[role, a public figure]. Topic: [topic].

Process it into:
1. QUOTES: pick 6 passages usable as direct quotes. For each, give the
   exact wording (don't edit the words, only remove obvious slips and
   filler and mark that you did), where it falls in the interview, and
   why it's strong (new claim / admission / number / phrasing).
2. EVASIVE SPOTS: where the speaker answered evasively, changed the
   subject, or answered a different question than the one I asked.
3. CLAIMS TO VERIFY: facts the speaker stated that I need to verify
   elsewhere before publishing them.
4. DIDN'T ASK: what's still open and should be followed up on.

Don't rephrase anything into nicer prose. The quotes have to be usable
as-is.

Transcript:
[paste the transcript]

Point 2 is something you easily miss while listening yourself, because at that moment you were thinking about your next question. Critical warning: this prompt is for a public figure who knows they're speaking on the record for the media. A transcript of an interview with a protected source doesn't go into any outside tool at all — why, and what to do instead, is in the section on protecting sources.

What from research is allowed into the piece

Structure, questions, lists of databases, and sorted material can be used directly — none of that is a claim about the world. Contradictions the model flagged get verified in the originals. And facts, numbers, dates, names, and quotes the model came up with on its own don't belong in the piece until you've seen them at the source. That includes phrasing that sounds right but you don't know where it came from — delete it. That's the most common way someone else's sentence ends up in an article.

Data from public sources

Public data on how the Czech state spends money is remarkably accessible — the Register of Contracts, public procurement, subsidies, political party donations, insolvencies, decisions from the competition authority. The problem was never availability; it's that searching it is slow, one entity at a time, click by click, and when you connect two things you end up retyping the numbers into your own spreadsheet. (Most countries have some equivalent of these registries, and the approach below carries over even if the names and tools differ.)

Hlídač státu (“State Watchdog,” a Czech public-sector transparency watchdog) collects, cleans, and connects this data — and has an MCP server, an interface an assistant can read directly. Instead of clicking through registries, you hold a conversation: a company name, a Czech company ID (IČO), a series of questions, connections drawn between them. Setting it up takes a few minutes and is covered in detail in vetting a business partner through Hlídač státu — that guide is written from a business angle, but the technical part applies without change; for connectors in general, see the overview of MCP connectors. The journalism angle gets its own piece: Hlídač státu for journalists.

The difference between a business vetting check and journalistic research

A business asks is this partner safe? — the answer is a traffic light, and that's it. A journalist asks is there something here the public doesn't know? — what interests them is a pattern, a repetition, a connection. One company on its own isn't a story.

Three differences follow from that. A journalist works with a time series, not a current snapshot — they don't care that a company exists today, but when it was formed and what it did before. They work with a set of entities, not one: a single contract means nothing, two hundred contracts from one authority over five years has a shape. And they work to disprove, not to confirm — you build a hypothesis in order to try to kill it.

What you can find in the data

A practical rundown of what's worth asking about:

Register of Contracts. Who signed what contract with whom, for how much, and when. Publication is mandatory for a large share of public procurement, so a missing contract in the register is itself a question.

Public procurement. Tender procedures, number of bids, winner, price. A tender with a single bid isn't illegal, but it's a place where competition didn't actually happen — a legitimate question for the contracting authority.

Subsidies. Who got money, from which program, how much, and for what. What's interesting is repetition (the same recipient across five rounds) and timing.

Political party donations. The one dataset that can connect “this company wins contracts” with “this company funds politics” — and also the easiest place to overreach. A donation isn't a crime or proof of anything; it's context.

Insolvencies and corporate criminal records. Who is or was insolvent, who's been penalized. Especially useful for contractors with a big contract and a thin track record.

Decisions from the competition authority (the Czech antitrust body). Who's already been slapped down over a tender procedure before; repetition is a story. And politicians and their income — asset disclosures and compensation, where they're in the data.

From a topic to a query

The most common mistake is asking too broadly. “Find me something on the region” won't return anything usable — a query needs a subject, a time period, and one question.

You have the Hlídač státu MCP server connected. I'm doing research for a
news piece, not a business vetting check — I'm interested in patterns
and anomalies, not a traffic light.

Subject: [name of the authority / municipality / organization]
Period: [from–to]
Question: who received money from this contracting authority, and how
did that change over time.

Proceed like this:
1. Find the subject's company ID (IČO) and confirm we have the right one
   (name, registered seat, legal form).
2. List contractors by total contract volume over the period —
   TOP 15, columns: contractor, IČO, number of contracts, total volume,
   first and last contract.
3. For the five largest, show the volume broken down by year.
4. Flag what's unusual in that data: a sudden jump, a contractor that
   appeared out of nowhere and straight to the top, repeated contracts
   just under a threshold, a long gap and a sudden return.
5. For each finding, note what could innocently explain it.

For every number, cite which record it's from, so I can open it myself.
Don't estimate anything — when data is missing, say it's missing.

Point 5 is there on purpose: it forces both the model and you to come up with the innocent explanation before you fall in love with a finding. Most “anomalies” in contract data have a boring explanation (a large capital project, a change in methodology, an organizational merger), and it's better to know that before you call the spokesperson.

Critical note: recheck any aggregated numbers yourself. Summing long lists is exactly where language models quietly get things wrong. When a total is going into the piece, download the data and add it up in a spreadsheet or a script.

The company on the other side

The second common direction: you have a contractor and want to know what they're about.

Using Hlídač státu, build me a data profile of the company [name],
IČO [12345678]. I'm writing a news piece about it, so I need facts and
dates, not assessments.

1. Basics: formation, registered seat, legal form, field, changes of
   seat and name.
2. People: statutory body and ownership structure, including changes
   and dates.
3. Money from the state: contracts with public institutions (year,
   contracting authority, subject, amount) and subsidies (program, year,
   amount, purpose).
4. Risk records: insolvency, corporate criminal records, unreliable
   VAT payer status, competition authority decisions.
5. Political ties in the data: donations to parties by the company or
   by people in its statutory body, with date and amount.
6. Timeline: every event found, chronologically, in a single list
   (date — event — source).

Point 6 matters most to me. If a category has nothing, write
"no record" — don't guess.

The timeline in point 6 is a trick almost nobody does with this data, and it's the one that shows the most: when “change of director,” “entered the tender,” and “political donation” sit next to each other with dates, you can tell right away whether they're connected in time. Verify the identification in point 1 — companies with similar names are the most common source of error, and IČO is the only solid ground.

Connecting the dots: where the story comes from

The third type of query is the reason you do any of this — finding the intersection between two sets.

Using Hlídač státu, test a hypothesis. I'd rather disprove it than
confirm it, so look for evidence against it too.

Hypothesis: companies that donated to party [party] between [from–to]
also won contracts from [contracting authority / type of authority].

1. List the companies that made a donation to the party in that period
   (company, IČO, date, amount).
2. For each one, find contracts and tenders with the contracting
   authority in the period [from–to] — count and volume.
3. Build a table: company | total donations | total contracts | first
   donation | first contract | which came first.
4. Separately, note which companies had contracts even before their
   first donation — that weakens the hypothesis.
5. Separately, note how many donors have no contracts at all.
6. Note what data would properly confirm or disprove this hypothesis,
   and which of it Hlídač doesn't have.

Don't draw conclusions about motive. Just return numbers and dates with
sources.

Points 4 and 5 are there so the piece isn't dishonest. If six out of fifty donors have contracts, and all six had them before donating, that's not a story about influence — it's a story about big companies both donating and doing business at the same time. Point 6 is an admission of the data's limits that belongs in the published piece too.

Where the data stops

Four things to keep in mind before you turn a database into a headline:

Data isn't proof of intent. It shows what happened, not why. Correlation in time isn't causation.

A missing record doesn't mean nothing happened — only that it's not in this particular database. Disclosure requirements have exceptions and thresholds.

A company with a similar name isn't the company. IČO, always IČO.

The other side gets a say. Before the piece runs, the people involved get specific questions with specific numbers and reasonable time to respond. That's not courtesy, it's part of the craft — and the fastest way to find out you mixed up two companies.

Writing and form

This is where AI saves the most time for the least risk, because nothing here decides what's true, only its shape. The condition is that the text already exists: the model doesn't edit your piece until it has one.

Headline and intro

You don't write a headline by generating one. You write fifteen and recognize which one is right. The model has no idea what the story actually means — but it's good at spreading variants across a range from dry to sharp.

This is a finished news piece. Suggest headlines.

Rules:
- No clickbait, no questions in the headline, no "shocking."
- The headline must be true even without the article — it can't promise
  more than the text delivers.
- Don't use words that aren't backed up in the text.
- Length up to [65] characters including spaces; give the character
  count for each.

Give me 12 variants split into three groups:
A) STRAIGHT NEWS — who, what, where. 4 variants.
B) LEADING WITH A NUMBER or a specific detail. 4 variants.
C) LEADING WITH THE CONSEQUENCE for the reader. 4 variants.

Then, separately, tell me which claim in the text is the strongest one
in your view, and why — I want to know if I gave it enough emphasis
in the piece.

Text:
[paste the article]

The last paragraph of the prompt is the most useful part: it often shows you that the most interesting thing is buried in the fourth paragraph. Check group C variants the hardest — that's exactly where models like to overstate the impact (“the region will lose millions” where the text says “the region risks losing money”).

The intro works the same way, just with different rules:

Write an intro for this piece. Three variants.

Rules:
- The intro can't just repeat the headline in different words.
- It must include: what happened, who it affects, and one specific
  detail (a number, date, or name) from the text.
- Length [280–320] characters including spaces; give the exact count
  for each variant.
- No adjectives that aren't backed up in the text.
- The intro's last sentence should be the reason to keep reading — but
  not a promise of something the text doesn't deliver.

Headline: [paste the headline]
Text:
[paste the article]

Cutting text to an exact character count

The most tedious job of the day. Models need an especially firm brief here: they're bad at counting characters and tend to cut specifics (numbers, names) while leaving general sentences alone, which is exactly backwards from what you want. A detailed process with a check step is in cutting text to an exact character count; here's the basic newsroom version.

Cut this text to [3800] characters including spaces. It's currently
[4600] characters.

What must NOT be lost:
- any number, date, name, or quote
- any claim that carries information
- the note that we contacted the other side

What to cut, in this order:
1. Repeating the same thing in different words
2. Explaining commonly known context
3. Subordinate clauses that only soften the phrasing
4. Modifiers that don't carry information
5. Only as a last resort: the least essential paragraph as a whole —
   and tell me which one and why

Don't cut in a way that chops sentences into fragments. The result has
to be a readable news text.

At the end, write:
- the result's length in characters
- a bulleted list of what you cut

Text:
[paste the text]

The model will almost certainly not hand you exactly 3,800 characters. Measure the result yourself and fine-tune the last few dozen by hand. The list of what got cut at the end is mandatory: it's the only way to notice that the sentence about the spokesperson declining to comment got lost along the way.

Tone

The same facts can be written for a news website, a print magazine, and a newsletter, and each one should sound different. Shifting tone is a mechanical task models handle well — the risk is that content shifts along with the tone. Covered in detail in changing a text's tone; from a newsroom angle, one rule belongs in the prompt every time: the facts don't change, only the language does.

Rewrite this text in a different tone. Facts, numbers, names, and
quotes must NOT change — not rounded, not simplified, not dropped.

Source tone: [official press release / trade text / spoken remarks]
Target tone: [straight news / narrative feature / newsletter written
in first person / an explainer for readers with no background in
the field]

Specifically:
- Sentence length: [short declarative sentences / longer compound
  sentences]
- Voice: [impersonal / newsroom in first person plural]
- Technical terms: [keep and explain in parentheses / replace with
  plain language]
- Banned phrasing: bureaucratic passive voice, throat-clearing
  officialese, nominalizations that hide the actor

At the end, list any changes that touched something other than style —
if you had to alter content for the sake of tone, I want to know.

Text:
[paste the text]

That last point is the check. Even when the model says “none,” read the result against the original with a pencil in hand — watch numbers especially, and modality (“may” versus “will”). Brand tone from a marketing angle is covered in brand tone of voice; a newsroom has an advantage over a brand in that its tone is already written down in a style guide, and you just have to paste it into the prompt.

Proofreading and typography

The last step before filing, and the one where the temptation to let the model “polish it a little” is the most dangerous. Proofreading means fixing spelling and leaving the text alone. How to phrase the prompt so the model doesn't touch style is covered in fix only the spelling; style-guide typography conventions are covered in typography conventions. For a newsroom it's worth adding a layer that ordinary proofreading skips:

Check this text before publication. DO NOT edit it — just list findings,
so I can decide myself.

1. SPELLING AND GRAMMAR: errors, subject-verb agreement, commas.
   Format: the flawed spot → the fix → reason (brief).
2. TYPOGRAPHY: quotation marks, dashes versus hyphens, spacing around
   abbreviations and numbers, ellipses, orphaned short words at line
   ends.
3. INTERNAL FACTUAL CONSISTENCY: numbers that repeat in the text and
   don't match; names spelled differently each time; dates that don't
   line up; totals that don't add up.
4. LEGALLY SENSITIVE SPOTS: sentences that assert a crime without
   citing a final judgment; an accusation phrased as fact; imprecise
   use of terms like "accused," "charged," "convicted," "suspected."
5. UNSUPPORTED CLAIMS: sentences with no source or link to where the
   information came from.

Don't rewrite anything. Just the list.

Text:
[paste the text]

Points 3 and 4 are why this prompt earns its keep over ordinary proofreading. A number that doesn't match between the intro and the third paragraph is the most common mistake in text written under deadline pressure, and you'll almost always miss it reading your own work. Point 4 is just a flag, not a legal opinion — that's done by a person, and for sensitive pieces, by a lawyer. For a tougher challenge than proofreading, there's AI as opponent: you have the model attack your own text from the other side's position. Before publishing an investigative piece, it's half an hour well spent.

Visualization and graphics

A visual in an article has one job: show something a sentence would say worse. When it doesn't do that, it's filler. Before every image, ask what exactly the reader will learn from it — and if the answer is “that someone was wearing a suit,” the image doesn't belong there.

When to chart, when to diagram, when to photograph

A chart belongs where there's a series of numbers and their ratio or trend matters: five years of contract volume, how subsidies break down, comparing two regions. A single number doesn't get charted — a single number gets written in bold.

A diagram or schematic belongs where the story is relationships, not quantities. Who sent money to whom, how a decision moves through levels of authority, who owns whom. In investigative pieces, this is the most useful image and also the one most often missing, because it's hard to draw.

A photo belongs where what something looks like, or who someone is, matters — and in news, a photo means an actual photograph of an actual place or person. Nothing is also a legitimate choice: an article about a methodological dispute between two agencies doesn't need a stock illustration of a computer.

From numbers to a chart

Before you turn a table into a chart, it's worth getting advice on the type — and above all, having the data's actual content spelled out, so the chart doesn't claim more than the data can support.

I have this data for an article about [topic]. I want to turn it into
one chart that will go in the piece.

Data:
[paste the table or list of numbers with labels]

1. State what's objectively in this data — what quantities, over what
   period, in what units, where the gaps are.
2. Suggest 3 chart types that would fit. For each, say what it
   emphasizes and what it hides.
3. Recommend one and explain why.
4. Give me: a chart title, axis labels including units, a data-source
   line for the caption, and a methodology note if one is needed.
5. Flag if any of those chart types would be misleading — a truncated
   axis, incomparable periods, mixing nominal and real values, a small
   sample.

Don't calculate any new values for me. Work only with what I gave you.

Point 5 is why it's worth doing this deliberately instead of eyeballing it: a truncated vertical axis is the most common way to turn boring data into a dramatic chart, and the most common reason an outlet ends up issuing a correction. Read point 1 carefully — if the model says the data covers different periods, you have a problem before the chart even exists. Don't leave rigorous calculation over a large dataset to a chat; the process for that is in reproducible data analysis.

Diagrams in SVG

A diagram of money flow or an ownership structure can take an hour to draw in a graphics editor, or you can describe it in words and have it generated as SVG in five minutes. SVG has three advantages for a newsroom: it's text, so it can be versioned and edited; it scales without quality loss for both web and print; and changing one number is a fix in the file instead of redrawing the whole thing. The process is in diagrams in SVG.

Create an SVG diagram to go in a news piece.

What it should show: [money flow from the contracting authority through
the main contractor to subcontractors / an ownership structure /
a timeline of events]

Elements and relationships:
[list: nodes and what connects them, including amounts and dates]

Requirements:
- 900px wide, height fits the content, readable even scaled down to
  600px.
- Sans-serif typeface, label size at least 13px.
- Restrained palette, three colors at most, no gradients or shadows.
- Arrows showing direction of flow, each labeled with amount and year.
- Diagram title top left, a data-source line bottom right.
- Must be legible in black and white too — distinguish by shape and
  line, not color alone.
- No text in the image that isn't in the data I gave you.

Return clean SVG code I can save as a file.

Always open the result and read it — models sometimes place a label over a line or truncate a long name; the fix is easy, since it's just text. And above all: check every number in the diagram against the source. A diagram reads as more authoritative than a paragraph, and a mistake in it gets picked up and repeated.

Images in Gemini, and where the line is

Generating images is trivial today, and in a newsroom it's also the most sensitive tool you have. A rule to tape above your desk:

You don't generate an illustration for a news story.

The reason isn't aesthetic, it's a matter of craft. A news context gives an image meaning. When a piece about a flood in a specific town runs with an image of a flooded street, the reader reads it as a photo from that town — regardless of what the caption says, because a minority of readers even read captions, and captions get lost anyway when the piece is reshared on social media. You've produced a documentary-looking frame of an event that never looked like that. No footnote fixes that, and it's exactly the kind of mistake that costs an outlet its credibility.

You can generate an abstract illustration for an explainer that can't pass for a photo (geometric, symbolic, obviously stylized), diagrams and schematics, graphic elements like backgrounds and fills, and internal previews for a designer that never get published.

You never generate anything that looks like a photograph of a real event, place, or person; a portrait of a real person, stylized or not; a reconstruction of “what it probably looked like”; or anything for a crime story, a tragedy, or a war.

And whatever does get generated gets labeled — visibly, next to the image, not buried in a masthead credit, even when it's obviously stylized. Gemini, ChatGPT, and other tools can generate images directly in chat; for repeated generation you can automate it via API, covered in generating images via the Gemini API.

Captions and alt text

A photo caption has to say who or what is in the image, when and where it was taken, and who the photographer is; alt text describes the image's content for a screen reader. The model can handle this on one condition: it has to see the image and can't guess at what's in it. Upload the photo, give it verified details (place, date, names), and let it assemble the caption — never the other way around. When the model writes a caption purely from the image, it'll add “pictured: the town's mayor” without actually knowing who that is. Only names you've verified belong in a caption.

From a finished piece to social media

The piece has run, and now it needs to become a Facebook post, something shorter for X, an Instagram card, and a line for the newsletter. Mechanical work, and the last spot in the day where you can save twenty minutes with no risk — on the condition that three rules hold: nothing that isn't in the article, no promise the text doesn't deliver on, and the link points to the article, not to a paraphrase.

A detailed process with templates for individual platforms is in turning an article into social posts; the marketing take on content production at scale is in content factory — but a newsroom holds itself to stricter rules.

Turn this published article into social posts.

Iron rules:
- No claim that isn't in the article. Not even a simplified one.
- No "guess what happened next" questions and no "shocking revelation."
- When the article says something only "risks" happening or is
  "according to the spokesperson," that qualifier has to survive into
  the post too.
- Copy numbers exactly, including time period and units.

Produce:
1. FACEBOOK: 3 variants, each [400–600] characters. One opens with a
   specific number, one with a human detail from the piece, one with a
   plain summary.
2. X: 3 variants up to [250] characters, with the character count for
   each.
3. INSTAGRAM: caption text for a card, [2–4] short sentences, plus a
   suggestion for the large text on the card itself (8 words max, must
   be a quote or a number from the piece).
4. NEWSLETTER: one sentence introducing the link, up to [180]
   characters.
5. Separately: a list of 3 sentences from the article usable as pull
   quotes for graphics — verbatim, with who said them.

At the end, note if anything you drafted claims more than the article
does.

Article:
[paste the published text]

That last line is a self-check the model surprisingly often uses against itself. Even so, read every post against the article. The most common drift: “the authority made a mistake, according to the opposition” becomes just “the authority made a mistake” on the way to social — a difference a lawyer will notice.

Deadlines and overview

The last operational layer. It's not about writing, it's about making sure nothing falls through the cracks and you know each morning what changed overnight.

Morning brief

A journalist has three sources each morning: their inbox, their calendar, and whatever happened overnight in their beat. Assembling one overview from that is a job for a scheduled task that runs before you get in. Setup, including connecting mail and calendar, is in a morning brief with Claude and routines over mail and calendar.

Put together my morning newsroom brief. Run this every weekday at 7:00.

Sources: my mail and calendar (via connectors) and the beats I follow:
[topic 1], [topic 2], [topic 3].

Structure:
1. TODAY: what's on my calendar, with times. Separately flag anything
   with a hard filing deadline.
2. AWAITING REPLY: who I wrote to and haven't heard back from — name,
   what it was about, how many days it's been pending. Sort by age.
3. NEW IN MY INBOX: press releases and invitations from the last 24
   hours, one line each, ranked by how closely they relate to my
   beats. Flag anything under embargo or with a deadline.
4. MY BEATS: what's new on them (with links and sources, no claims
   without a link).
5. ONE THING: what you think I should do first today, and why.

Be concise. The whole brief should fit on one screen. Don't infer
anything — if something isn't in my mail or calendar, say it isn't
there.

Point 2 pays off the most: an unanswered question to a spokesperson is the most common way a story gets lost in a newsroom. Check point 4 against the actual links — a brief with claims and no source is worse than no brief at all, because it's easy to start quoting from it.

Tracking deadlines

The filing deadline isn't the only deadline a journalist tracks: there are embargoes, institutions' response windows, statutory deadlines under freedom-of-information law, scheduled calls. Estimating how much fits in a day tends to run systematically optimistic — the method is in a realistic deadline, and tracking unanswered questions is in the follow-up tracker.

The practical minimum: log every question you send an institution with a date and its response window, and set a reminder three days before it lapses. Without that, you end up publishing “the authority didn't respond to our questions” while its window is technically still open — and that's a mistake on your side.

Connectors

Beyond mail and calendar, you can connect a document store, a note-taking system, a newsroom wiki, or your own database; an overview is in MCP connectors and tools. Three directions are especially interesting for a newsroom: an archive of your own past pieces (so you can find what you wrote about a topic three years ago), a shared beat database, and public data like Hlídač státu. The same question applies to every connector: what exactly am I giving the assistant access to? A connector to your entire company drive means the model can also see the folder with investigations still in progress.

Protecting sources and ethics

This section is the reason a journalism guide can't be written like a marketing one. Everything else in this piece is about efficiency. This is about the fact that getting it wrong hurts a real person and ends your career.

What never gets uploaded

Write it on paper and stick it to your monitor:

A protected source's name, contact details, or any identifying description. Never, into any tool, not even a paid one, not even “just for a moment to sort it out.” That includes a phone number, an address, a workplace name, a role at a small organization, and photographs.

Unpublished material you received from a source. Internal documents, emails, spreadsheets, photographs of documents — even when no name appears in them. For a document only three people have seen, a leak of its content is the same as a leak of the name.

A transcript of an interview with a protected source. Even one that looks anonymized at first glance: speech is full of details that identify a person — where they worked, when they started, how many people report to them.

Material from an ongoing investigation under a newsroom embargo, and contact lists — a source book is the single most valuable thing a journalist owns, and it has no business anywhere outside your own device.

The difference between a paid account and an anonymous chat

That difference is real, but it can't be used to draw the wrong conclusion.

What a paid business or team plan changes: you can sign a data processing agreement under data-protection rules, content doesn't feed into training by default, the newsroom has central account management and can turn accounts off when someone leaves, there's an audit trail with a controlled retention period, and which connectors are allowed can be set centrally. A free or personal account reliably gives you none of that: data handling tends to be set up differently, you can't sign an agreement, and you have no control over what happens to the content afterward.

And what even the paid version doesn't change: it's still someone else's system, one the provider can access and that can be reached through legal process. Contractual data protection guards against commercial misuse and negligence. It is not protection for a source against law enforcement, and it is not journalist-source privilege. You protect a source facing dismissal or prosecution by making sure their identity doesn't exist anywhere in digital form outside your own head and your secured notes — not by paying for a better plan.

The practical upshot: a business plan makes sense for a newsroom, because everything routine then runs in an environment with clear rules. And at the same time, it's worth having a category of material that goes into no tool at all — and everyone in the newsroom needs to know it, not just whoever's working the investigation.

Anonymizing material when you have to process it

Sometimes you have fifty pages of documents and need them sorted or you'll never make deadline. There's a way to do it, but the order can't be skipped: clean first, then upload. Not the other way around, and not “I'll upload it and just tell it to ignore that part.”

A person does the cleaning, not the model — a model you hand material to for anonymizing has already seen that material. Go through the document and replace people's names with roles (Person A, department head), workplace names with generic ones, specific dates with relative ones, round off amounts, and strip out phrasing characteristic of the author. Only after that does it make sense to ask whether the cleanup was thorough enough:

This is an anonymized working document. I want you to check the
anonymization, not the content.

Go through the text and find spots that could reveal the identity of a
specific person or organization, even without a name present:
1. A unique combination of details (role plus time period plus
   organization size, a specific project, a specific event).
2. Details that are harmless on their own but together narrow the field
   down to a handful of people.
3. Phrasing that reveals a relationship to the matter (who had access,
   who was present, who could have known).
4. Any leftover specific numbers, dates, or names I missed.

For each finding, say how to rewrite it so the information stays usable
but the trail disappears.

Don't summarize the content for me, don't analyze it, don't draw any
conclusions from it.

Text:
[paste the cleaned material]

Use this prompt only on material you've already cleaned — it's a second check, not the first step. And even after it: if exposing the source would carry serious consequences for that person, the material doesn't go into an outside tool at all, cleaned or not. The level of risk to the source decides, not your convenience.

Transparency with the reader

The rule fits in one sentence: you label what touched the content, not what touched the process.

Label a generated image, a machine translation, an automatically produced transcript published as a transcript, an automatically generated summary at the top of a piece, and any text that a human didn't write or check. You don't have to label the ordinary tools of the trade: spell-checking, headline suggestions an editor chose from, an approved trim, research support, sorting material. Nobody discloses that a piece went through spell-check in their editor, and AI proofreading is the same thing.

The gray area between those gets settled with a newsroom policy, not case by case. One page on what AI is and isn't allowed to do at your outlet, and what gets labeled, saves a lot of arguing — and gives you something to point to when someone asks.

Authorial responsibility

When a mistake makes it into an article, “AI wrote that” isn't an answer. It's not an excuse or a mitigating factor — it's an admission that you didn't check the text. Responsibility is undivided and non-transferable: the named person stands behind every sentence, number, quote, and image.

That gives you an operational rule worth holding onto even under pressure: before a piece goes out, read the whole thing, slowly, start to finish. Not skimming, and not just the parts you edited. Most of the mistakes that make it into AI-assisted text aren't dramatic hallucinations — they're a shifted number, a lost qualifier, a vanished sentence about contacting the other side. All of them turn up on a slow read, and none of them on a fast skim.

Klára's day, after

A month into this approach, her schedule looks like this: press releases from 40 down to 15 minutes (the comparison prompt, then targeted reading of the originals), digging from 50 down to 25 (data through the connected Hlídač státu, totals rechecked in a spreadsheet), trimming from 25 down to 10, transcribing the interview from 60 down to 20, and social from 20 down to 7.

She doesn't spend the roughly ninety minutes she saves on a third piece. She spends it on two things: calling the other side, and reading her text slowly before sending it. She used to do both only sometimes. What hasn't changed: she's seen every number at the source, she has every quote on the recording, and her name is on the piece.

Common mistakes

• Using a fact from the model without verifying it, because it sounded credible. The most common and most expensive mistake. Confident phrasing isn't a sign of truth. Whenever a reply doesn't have a link you can click, it isn't information.
• Uploading source material “just for a moment.” A document once uploaded can't be taken back. Deadline pressure is exactly when this happens — which is why the rule has to be simple enough to hold even at five past midnight.
• Letting the model add up a total and dropping it into the text. Models quietly get arithmetic wrong over long lists and give no sign of it. Numbers that go into an article get calculated in a spreadsheet or a script.
• Trimming a text without checking what got cut. The model cuts what looks like padding — and “the spokesperson didn't respond by deadline” looks like padding. Without a list of what was cut, you won't notice.
• Generating an illustration for a news story. In a news context, the reader reads the image as a photo from the scene. A caption doesn't fix it, because captions vanish when the piece gets reshared on social media.
• Letting the meaning shift on the way to social media. “According to the opposition, the authority erred” becomes “the authority erred.” The article is fine; the post is actionable.
• Assuming a paid account protects a source. It protects data against commercial misuse. It isn't privilege, and it won't hold up against legal pressure.

The best tools

• Claude (claude.ai) — the main workspace for research, sorting material, working on form, and connecting connectors. Projects are useful for beats you follow long-term: upload your own style guide, past pieces, and notes, and you don't have to repeat context in every chat.
• The Hlídač státu MCP server — public data on contracts, deals, subsidies, donations, and insolvencies, right in the conversation. The journalism workflow is covered in Hlídač státu for journalists, the technical setup in vetting a business partner.
• A tool with search and citations — for anything that needs to rest on a source. Without search turned on, you're working from the model's memory, which is unusable for news.
• Deep research — a structured overview with links, when it's worth letting the model work for longer. Good for background on a topic, not for breaking news.
• NotebookLM — answers only from what you upload to it. Useful for large sets of public documents (council minutes, annual reports) where you want certainty the answer didn't come from anywhere else. The same sensitivity rules apply there.
• A dictation or transcription tool — for transcribing recordings. Fine for public appearances without reservation; never for protected sources.
• An editor with a character counter and a spreadsheet — the check after every trim and every total that goes into the piece. The model only estimates both.

What you get out of it

Time. Realistically an hour or two a day for a reporter filing two pieces daily — and it's time stolen back from the mechanics, not from the writing. The biggest single items: transcribing interviews, sorting press releases, trimming, and social.

Money. Directly for a freelancer: the same hours worked, more pieces filed, or the same number of pieces and time to make one of them better. Indirectly for a newsroom: the time saved can go into verification, which is the one investment that keeps an outlet's credibility over the long run.

Peace of mind. An unanswered question to an authority doesn't get lost, because a routine tracks it. The deadline stops being a race against trimming. You know each morning what changed overnight before you even open your inbox.

Quality. Honestly: AI by itself doesn't raise the quality of your writing. What you do with the time you save does. Spend it on a second phone call and a slow read before sending, and quality goes up meaningfully. Spend it on a third piece a day, and it goes down.

Pro tip

Set up a project (persistent context) and upload three things to it: your newsroom's style guide, ten of your own pieces you consider your best, and a one-page rule for what never goes into the tool. From then on you don't have to explain tone and rules in every prompt — and, most importantly, that third rule is in front of you every time you're about to paste something into a chat.

And one last thing that sums up this whole piece. Before you send anything, ask yourself the question a reader would ask if they could: how do you know that? If the answer is “I found it in a database, here's the record,” “a named person told me, I have it on tape,” or “it's in a document I have,” you're fine. If the answer is “the chat gave it to me,” you don't have an article. You have a hypothesis that needs verifying.