AI in the Newsroom: What 27 Pakistani Journalists Told Me

I did not teach a newsroom to trust AI. I taught it to interrogate AI. That distinction is the entire method, and it is why I now believe a full master class is no longer optional.

This is a field report from an AI in the Newsroom workshop I delivered to the Web Desk and Social Desk of the Ausaf Group, a Pakistani media house running across print, broadcast and digital. It covers who was in the room, what a short survey of 27 participating journalists revealed about which of them are readiest to adopt AI, and — published here in full — the complete prompt system I left the desk with, so any journalist reading this can copy it and run it today.

Key takeaways

• 96.3 percent of the 27 journalists surveyed wanted a follow-up AI in Journalism Master Class. Twenty-six out of twenty-seven said yes.
• Appetite did not fall among the least confident. It rose. Of the five journalists who rated their own likelihood of using these tools at the bottom of the scale, 100 percent still wanted the master class — against 95.5 percent of everyone else.
• Newer journalists cluster on “maybe.” Experienced journalists split. Among the 1–5-year cohort, 47.8 percent picked the exact midpoint of the scale. In the 6–10-year cohort, not one person did — they landed at the top or the bottom, nowhere in between.
• 50 percent of the 6–10-year cohort rated their adoption intent at 4 or 5, against 34.8 percent of the 1–5-year cohort. With only four journalists in the senior band, this is a signal to test, not a settled finding.
• The complete master prompt system is reproduced in full below: a prompt that turns documents into cited drafts, three that scan a live beat every morning, one that builds a house style guide from a newsroom’s own published work, and one that converts a recurring page into an interactive dashboard.

What was the training, and who was in the room?

The workshop went to the digital teams: the Web Desk and the Social Desk. The room held the mix I now expect in every Pakistani newsroom — young journalists a year or two into the job sitting beside careered reporters and desk editors with a decade behind them. Both groups arrived with the same private question, phrased two different ways. The younger ones asked how to use AI without getting caught out. The senior ones asked how to use it without lowering their standards. Those are the same question. The answer is a method, not a tool.

The method has a name I use out loud in the room: interrogate the machine. A journalist already knows how to interrogate a source. You check who is speaking, what they gain, what they left out, and whether a second source confirms it. AI output deserves precisely that treatment and almost never receives it, because it arrives polished, fluent and instant, and fluency reads as authority. It is not authority. It is formatting.

The single most valuable thing a newsroom can install is not software. It is the reflex to distrust a clean answer until the fact inside it has been traced back to a real document.

Why does interrogating AI matter more than any tool?

Because your audience has already reached a verdict, and it is not a kind one. In the Reuters Institute’s 2025 Generative AI and News report, only about 33 percent of people believed newsrooms “always” or “often” verify AI outputs before publication. Read that from the desk’s side. Two-thirds of the public assume that when a newsroom uses AI, nobody checks its work.

That is a trust problem, and trust is the one asset a newsroom cannot buy back once it is spent.

Here my strategy practice and my newsroom training converge. My positioning as a strategist is listen first — audience intelligence before spend. In a newsroom the translation is nearly word-for-word: read the source for what it is before you write a single line. I build to a Trust-Led Communications Framework, which holds that every communication decision is judged first by whether it protects or spends the audience’s trust. A system that produces ten stories an hour is worthless if one carries a fabricated quote, because a single correction costs more trust than nine correct stories earn.

So the design goal was never speed on its own. It was speed with the checks welded in, so that the fast path and the safe path are the same path.

What did the participant survey measure?

At the close of the session I put two plain questions to the room through a short form, alongside how long each person had worked in journalism:

1. How likely are you to use Gemini Notebook, Deep Research, Canvas and Gems in your workflow? — answered on a 1-to-5 scale, where 5 is most likely.
2. Are you interested in attending an AI in Journalism Master Class? — yes or no.

Twenty-seven journalists gave valid responses. 85.2 percent (23) were in their first five years; 14.8 percent (4) were in their sixth to tenth year. This is a small, self-selected sample from one training, and it describes this room rather than the profession. I report it as field data, not as representative research, and I would ask any reader to hold it at exactly that weight. What follows is an honest reading, including the places where the numbers are too thin to carry the claim.

Who in the newsroom is most ready to adopt AI?

Here is the full distribution, by experience band, in percentages.

Adoption intent (1–5)	1–5 years (n=23)	6–10 years (n=4)
1 — least likely	8.7%	25.0%
2	8.7%	0.0%
3 — midpoint	47.8%	25.0%
4	21.7%	25.0%
5 — most likely	13.0%	25.0%
Rated 4 or 5 (high intent)	34.8%	50.0%
Mean score	3.22	3.25

The means are almost identical, 3.22 against 3.25. On averages alone there is no story here, and a less careful analyst would stop and report none.

The story is in the shape of the responses, not their average.

Newer journalists cluster on “maybe.” Nearly half of the 1–5-year cohort — 47.8 percent — chose the exact midpoint of the scale. That is the answer of someone who has heard enough about AI to know they should say something positive, and not enough to commit. It is curiosity without a plan.

Experienced journalists refuse the middle. In the 6–10-year band the distribution splits clean: 25 percent at the very bottom, 25 percent at the midpoint, 25 percent at 4, and 25 percent at the top. Half the cohort landed at 4 or 5 — 50 percent versus 34.8 percent for the newer group. But note what is missing. Not one senior journalist chose a 2. They either saw where this fits their workflow and committed, or they had already decided against it.

My working read, from the room and the numbers together, is that experienced desk journalists carry a finished mental map of the job. A new tool either slots into that map or it does not, and they can tell within minutes which it is. Their “yes” is concrete rather than curious — and so is their “no.” Newer journalists have no map yet, so they hedge.

I will not oversell this and neither should anyone repeating it. There were four journalists in the 6–10-year band. A pattern built on four people is a hypothesis, not a finding. To prove it, I would run these same two questions across three or four newsrooms and several hundred journalists before putting it anywhere near a headline. The honest version: in this room, seniority tracked with polarised, decisive intent, while inexperience tracked with hedging. Worth testing at scale. Not yet worth asserting.

Naming that limit is not a hole in the analysis. It is the analysis.

Do the least confident journalists want to opt out?

This is the finding I trust, because these numbers are not thin.

Interest in a follow-up master class was near-universal: 26 of 27, or 96.3 percent, said yes. One person said no. By experience band, 95.7 percent of the 1–5-year cohort and 100 percent of the 6–10-year cohort wanted it.

The sharper point is where that interest sat. I isolated the journalists who rated their own likelihood of using the tools at the very bottom, a 1 or a 2. There were five of them.

Group	Wanted the master class
Rated own AI-use intent lowest (1–2), n=5	100%
Everyone else (3–5), n=22	95.5%
All respondents, n=27	96.3%

Read plainly: low confidence with the tools did not produce low interest in learning them. The least confident group was the only group with unanimous interest.

This matters for anyone planning newsroom training, because the instinct runs the other way. The instinct is to design for the enthusiasts, the people already scoring themselves a 5. The data says design for the hesitant. The journalists who most need structured training are asking for it by name, and they are precisely the ones a one-off demonstration leaves behind. A demo excites the confident. A written method rescues the hesitant.

That is why I hand a newsroom a copy-and-paste system rather than a slideshow. A hesitant journalist can take a written system home, run it once, and get a correct result with no trainer in the room. A slideshow cannot do that. Which brings me to the system itself, published here in full.

The Newsroom Master Prompts

Four tools. The first turns documents you already hold into a checked story. The second goes out and finds the day’s stories on your beat. The third writes the style guide the newsroom never got round to writing. The fourth rebuilds a recurring page into something a reader can actually use.

Tool one: turn any source into a checked draft, using Gemini Notebook

Gemini Notebook — the tool formerly called NotebookLM — answers only from material you give it, and attaches a clickable citation to every claim that opens the exact line in your source. That constraint is what makes it safe for newsroom work where a general chatbot is not. It cannot wander off and invent a fact, because it has nothing to invent from.

This prompt turns any source — a price table, a policy document, a podcast, a press release — into three story angles, the calls you need to make, and a rough first draft with headline options.

Copy everything inside the box.

You are a senior news editor at a Pakistani newsroom. Use ONLY the sources in this notebook. Add no outside facts. Attach a citation to every fact, figure and quote you use.

STEP 0 — READ THE MATERIAL FOR WHAT IT IS
In one line, say what kind of source this is (data table, policy or legal document, report or study, speech, interview or podcast, press release, published news, or a mix) and who produced it. Then apply the matching lens:
- Numbers or data: what changed, what is highest and lowest, what is an outlier, and what the numbers cannot tell us. Show the basis of any calculation you make.
- Policy, law or report: what is promised, by when, by whom; what is left undefined; where the text contradicts itself or another source.
- Speech, interview or podcast: separate verifiable claims from opinion. Quote exactly, with the timestamp or location if the source gives one. Flag anything new, surprising or disputable.
- Press release or announcement: what is not said, who benefits, and what a sceptical reader would ask.

STEP 1 — THREE ANGLES
Give three genuinely different story angles, each a different type: (a) hard news — the most important new fact; (b) people — who is affected and how; (c) explainer — what a reader needs in order to understand this. For each: a one-line description, the key fact with citation, who it matters to, and why it is news now.

STEP 2 — PICK ONE
Name the strongest angle and justify it in two sentences.

STEP 3 — WHERE HUMAN REPORTING IS NEEDED
For the chosen angle, list every point that needs an outside voice. For each: the claim, WHO to ask (a role, or a named person if the source names one), and the exact question. If anyone or any organisation is criticised, include their right of reply as one of the calls.

STEP 4 — ROUGH DRAFT AND HEADLINES
Write a rough news story on the chosen angle, under 1000 words, plain language, for a general reader.
- First sentence carries the most important finding.
- Every fact traceable to a source; attribute to the source by name. Mark each place for an outside quote as [QUOTE NEEDED: who]. Never invent a quote.
- Report, don't judge: no opinion, no speculation about motive, no blame. On contested matters, present each position as the sources state it.
- Then six headlines, all under 12 words, no question marks: two plain and factual, two built on the key number or fact, two built on the central tension.

STEP 5 — VERIFY BEFORE PUBLISHING
List the three facts in the draft most in need of checking against the original source, and how to check each. If the material does not support three distinct angles, or is not newsworthy, say so plainly and explain why.

How to run it, step by step

1. Open Gemini Notebook. On a computer, go to notebook.google.com and sign in with a Google account. On a phone, install the Gemini Notebook app — it may still be listed under its old name, NotebookLM — and sign in with the same account.
2. Create a new notebook. Choose Create new. Give it a name that tells you what is inside: the story, the beat, or the date.
3. Add your sources. Choose Add source. You can add a website link, a YouTube video, a PDF, an audio file, or text you paste in. On a phone there is no file upload, so use links and pasted text. Add everything belonging to the story; three to six sources is a good number. Wait for each to finish processing before continuing.
4. Set your output language if you work in Urdu. Open Settings, then Output language, and choose Urdu. This governs the chat replies and the audio and other outputs. Skip it and everything returns in English.
5. Paste the prompt. Tap into the chat box at the bottom, paste the whole prompt exactly as written above, and send.
6. Read the output in order. The notebook works through the five steps. Read the three angles first. If you would rather build a different angle than the one it picked, reply with one line — “Draft angle (b) instead” — and it rewrites Steps 3 and 4 on that angle.
7. Click a citation. Every fact comes with a small numbered citation. Tap one and it jumps to the exact line in the original source. Do this every time for the number that ends up in your headline.
8. Make the calls. The draft carries markers reading [QUOTE NEEDED: who]. Those are the gaps the document cannot fill. That is where your reporting goes — the notebook has already told you who to call and what to ask.
9. Verify, then publish. Step 5 lists the three facts most in need of checking. Check them against the source before anything goes live. The prompt gets you a draft in ninety seconds. A human still signs it.

What you get back

• A one-line description of what the source is and who produced it
• Three different story angles, each with its key fact and citation
• The strongest angle, with the reason
• A list of people to call, with the exact question for each
• A rough draft under 1,000 words, with quote markers where reporting is needed
• Six headline options
• Three facts to verify before publishing

Tool two: find your beat’s stories every morning, using the Gemini app

The Notebook prompt turns documents you already have into a story. This one does the opposite job. It goes out and findsthe stories on your beat, live, from across the web, so a reporter or desk editor can walk in, run one prompt, and have the morning’s most newsworthy items ranked and ready in under three minutes.

Three ready-made versions follow — Technology and AI, Education and Scholarships, and Showbiz and Entertainment — each pointed at the leading international and Pakistani sources for that beat.

The one setting that makes this work

This runs in the Gemini app, not Gemini Notebook, because only the main Gemini app can search the live web.

1. Open the Gemini app on your phone or at gemini.google.com.
2. Start a new chat.
3. Turn on the tool that lets Gemini search the web. Depending on your version it is labelled Search, Grounding with Google Search, or shown as a small globe or Deep Research. Skip this step and Gemini answers from memory instead of today’s news, and the results come back old or wrong.
4. Paste the prompt for your beat and send it.

Turn search on, paste your beat’s prompt, read the ranked stories, pick one, verify the key fact, file. That is the whole routine.

How to read what comes back

Each prompt returns the same five things, in order:

• Top stories — the most newsworthy items from the last 24 hours, each with a source, a link, how recent it is, and a relevance score out of 5 for a Pakistani reader.
• Already everywhere / deadline alert / verified vs rumour — a check that helps you find fresh ground rather than chasing what every outlet already has.
• Worth covering today — the three strongest stories to write now, and why.
• The angle — for the single best story, a Pakistan-focused angle and who to call to stand it up.
• Verify — the facts to confirm against the original source before publishing, with links.

Beat prompt: Technology and AI

You are a news editor's research assistant for a Pakistani newsroom. Search the live web now. Grounding with Google Search must be ON.

BEAT: Technology and Artificial Intelligence
MARKET: Pakistan — for a Pakistani news audience.

PRIORITY SOURCES — check these first, then widen the search beyond them so you don't miss a scoop elsewhere:
- International: TechCrunch (techcrunch.com), The Verge (theverge.com), MIT Technology Review (technologyreview.com)
- Pakistan: ProPakistani (propakistani.pk), TechJuice (techjuice.pk), Dawn Technology (dawn.com/tech)

Find the most newsworthy stories on this beat from the LAST 24 HOURS — both Pakistan-specific developments and major global tech/AI stories a Pakistani audience would care about (new AI models and tools, big product launches, telecom and PTA news, IT exports, startups, funding, data and privacy).

Return, in this order:

1. TOP STORIES — up to 8, most important first. For each:
   - one-line headline
   - one sentence on what happened
   - the outlet and a working link
   - how many hours ago (skip anything older than 24 hours)
   - RELEVANCE 1-5 for a Pakistani reader (5 = highest), one-word why

2. ALREADY EVERYWHERE? — for each top story, say whether Pakistani media has already widely covered it (Yes / Not yet / Unsure). This tells me where the fresh ground is.

3. WORTH COVERING TODAY — the 3 strongest for us to write now, and why: fresh, relevant, and not yet saturated in Pakistani media.

4. THE ANGLE — for the single best story: a Pakistan-focused angle, the one fact that makes it a story, and who I would call to stand it up.

5. VERIFY — any figure, date, name or claim above I must confirm against the original source before publishing, with the link to check it against.

Use only what you find in your live search. If you cannot confirm a story is real and current, say so rather than including it. Never invent a source, a link, a quote or a number.

Beat prompt: Education and Scholarships

You are a news editor's research assistant for a Pakistani newsroom. Search the live web now. Grounding with Google Search must be ON.

BEAT: Education and Scholarships
MARKET: Pakistan — for a Pakistani news audience, especially students and parents.

PRIORITY SOURCES — check these first, then widen the search beyond them so you don't miss a scoop elsewhere:
- International: Times Higher Education (timeshighereducation.com), Scholars4Dev (scholars4dev.com), Scholarships360 (scholarships360.org)
- Pakistan: ilmkidunya (ilmkidunya.com), IlmiWorld (ilmiworld.com), Dawn Education (dawn.com — Education section)

Find the most newsworthy stories on this beat from the LAST 24-48 HOURS — both Pakistan-specific and major international items a Pakistani student would care about (newly opened scholarships and deadlines, HEC news, admissions, MDCAT/ECAT, exam and results updates, study-abroad and visa changes, fully funded international scholarships open to Pakistanis).

Return, in this order:

1. TOP STORIES — up to 8, most important and most time-sensitive first. For each:
   - one-line headline
   - one sentence on what it is
   - the outlet and a working link
   - the deadline or date if it is a scholarship or admission
   - how recent it is
   - RELEVANCE 1-5 for a Pakistani student (5 = highest), one-word why

2. DEADLINE ALERT — list any scholarship or admission closing within the next 14 days, soonest first. These are the most urgent to publish.

3. WORTH COVERING TODAY — the 3 strongest for us to write now, and why.

4. THE ANGLE — for the single best story: a Pakistan-focused, student-useful angle (who is eligible, how to apply, what it's worth), and which office to call to confirm.

5. VERIFY — any deadline, amount, eligibility rule or figure I must confirm against the official source before publishing, with the link. Scholarship deadlines and amounts change — flag every one.

Use only what you find in your live search. If you cannot confirm a scholarship or deadline is real and current, say so rather than including it. Never invent a source, a link, a deadline or a number.

Beat prompt: Showbiz and Entertainment

You are a news editor's research assistant for a Pakistani newsroom. Search the live web now. Grounding with Google Search must be ON.

BEAT: Showbiz and Entertainment
MARKET: Pakistan — for a Pakistani news audience.

PRIORITY SOURCES — check these first, then widen the search beyond them so you don't miss a scoop elsewhere:
- International: Variety (variety.com), The Hollywood Reporter (hollywoodreporter.com), Deadline (deadline.com)
- Pakistan: Dawn Images (images.dawn.com), The Express Tribune Life & Style (tribune.com.pk), Something Haute (somethinghaute.com)

Find the most newsworthy stories on this beat from the LAST 24 HOURS — lead with Pakistani showbiz (Lollywood film and drama, music releases, TV, awards, festivals, notable celebrity news) and include major international entertainment a Pakistani audience follows.

Return, in this order:

1. TOP STORIES — up to 8, most newsworthy first. For each:
   - one-line headline
   - one sentence on what happened
   - the outlet and a working link
   - how many hours ago
   - RELEVANCE 1-5 for a Pakistani reader (5 = highest), one-word why

2. VERIFIED vs RUMOUR — mark each story CONFIRMED (an official announcement, statement or credible outlet) or UNVERIFIED (social media claim, gossip, single unnamed source). Do not present an unverified item as fact.

3. WORTH COVERING TODAY — the 3 strongest confirmed stories to write now, and why.

4. THE ANGLE — for the single best story: the angle, the confirmed fact behind it, and whose statement or representative to get.

5. VERIFY — anything above I must confirm before publishing, with the link, plus a note on any story that is currently only a social-media claim.

Use only what you find in your live search. Do not repeat gossip as fact, do not invent quotes, and never fabricate a source or link. If a story is only an unconfirmed rumour, label it clearly as such.

How do I adapt these prompts to my own beat?

The three above cover common desks, but the structure works for any of them. To build your own — business, sports, health, city news — copy any beat prompt and change exactly two things:

• The BEAT line. Name your beat and, in brackets, the sub-topics you care about.
• The PRIORITY SOURCES. Swap in the two or three international and two or three Pakistani sites you trust most on that beat.

Change nothing else. The structure, the relevance scoring and the verify step stay as they are.

The Notebook prompt takes the same treatment. Change “a Pakistani newsroom” to your outlet’s name and audience. Change “under 1000 words” to whatever your format needs. Add a line to Step 4 carrying your house rules — no exclamation marks, always give the source’s full name on first mention, whatever your style guide says. A prompt is where a desk writes its standards down once instead of repeating them in every brief.

Save your version somewhere the whole desk can reach it, so every reporter runs the same disciplined routine.

Tool three: build the style guide the newsroom never wrote

This newsroom had no written style guide.

That is not unusual and it is not a criticism. Most Pakistani newsrooms run on an unwritten style carried in the heads of three or four senior people. It works until the desk grows, or those people are on leave, or a new sub-editor joins on a Tuesday and has to guess. Then the same designation is written three ways in one bulletin, and the outlet looks careless to the only people whose opinion counts.

The usual fix is to lock a committee in a room for a month. We did it in an afternoon, and we did it the honest way round: instead of writing down the style the newsroom thinks it has, we had Gemini read what the outlet had actually published and describe the style it does have. A style guide inferred from the evidence beats one invented from memory, because the desk recognises itself in it and therefore follows it.

Then we turned that guide into a Gem — a saved, reusable assistant in Gemini that carries the same instructions every time, so every sub-editor on the desk is working to the identical standard without anyone having to remember it.

Step one: have Gemini read the outlet and write the guide

1. Open your outlet’s website in one browser tab and copy the address from the bar at the top.
2. Open a second tab and go to gemini.google.com, or open the Gemini app.
3. Turn on live web search — labelled Search, Grounding with Google Search, a small globe, or Deep Research, depending on your version. Without this, Gemini cannot read your site and will invent a guide instead of finding one.
4. Paste the prompt below, with your outlet’s address dropped into the first line, and send it.

You are a senior news editor and style-guide author. I am giving you the website of a news outlet. Read as many published articles from it as you can reach — across news, business, sport, entertainment and any other section available — and infer the outlet's actual house style from what it publishes, not from what a textbook would recommend.

WEBSITE: [paste your outlet's web address here]
LANGUAGE OF THE OUTLET: [English / Urdu / both]

Read widely before you write anything. Prefer recent articles. Cover at least three different sections, and include both short news items and longer pieces.

Then produce a complete house style guide covering:

1. IDENTITY — the outlet's name, its sections, its languages, and who its reader appears to be judged from the coverage itself.
2. STRUCTURE — how a story opens, how long the first paragraph runs, where the most important fact sits, typical story length, how paragraphs are broken, how stories end.
3. HEADLINES — typical length in words, capitalisation, whether questions or wordplay are used, whether numbers appear, whether the section or outlet name is used.
4. ATTRIBUTION — how sources are named on first and later mention, how titles and designations are written, how wire and agency copy is credited, how quotes are introduced and punctuated.
5. NUMBERS, DATES AND UNITS — how figures, currency, percentages, dates and times are written, including local conventions such as lakh and crore if the outlet uses them.
6. NAMES AND PLACES — how Pakistani and foreign names are spelled and ordered, how cities and provinces are written, honorifics, and any recurring house spellings.
7. LANGUAGE — sentence length, formality, use of the active voice, words the outlet clearly favours, and constructions it clearly avoids.
8. WHAT THIS OUTLET NEVER DOES — anything it consistently avoids in tone, framing or presentation.

Then add these two sections, written as direct instructions to the assistant that will use this guide:

9. COMPLETENESS CHECK — before returning any story, examine the material supplied to you and list, under a clear heading "MISSING", every element a complete story of this type requires that the supplied material does not contain. Check at minimum for: a named human source or direct quote; a second independent source; the key number or statistic and where it came from; the date and location of the event; the full name and designation of anyone mentioned; a right of reply for anyone criticised; and the original source of any wire or agency copy. Name each gap plainly, say who to ask or where to look to fill it, and never invent, estimate or infer a missing element. If nothing is missing, say so in one line.

10. HARD RULES — never invent a quote, a name, a figure or a source; never strip attribution to a wire service or another outlet; never change the meaning of a quote while restyling it; report, do not judge.

Write the guide as plain numbered instructions in the second person, addressed to the assistant that will use it. Keep it under 1,200 words so it fits inside a Gem instruction field. Give real examples drawn from the articles you actually read, and name the article each example came from.

If the site blocks you, or you can only reach a small number of articles, say so plainly and tell me how many you were able to read. Do not invent a style guide.

Read what comes back before you accept it. If it describes a newsroom you do not recognise, tell Gemini which parts are wrong and have it correct them. The guide is a draft of your own practice, not a verdict on it.

Step two: turn the guide into a Gem

A Gem is a saved assistant inside Gemini that keeps the same instructions permanently, so you do not paste the style guide again every time.

1. Copy the whole style guide Gemini produced.
2. Open the Gem manager. In the Gemini web version, look at the left-hand menu for Gems — or click the settings wheel and choose Gems / Manage Gems. In the app, open the menu and find the same option.
3. Click New Gem (it may read Create a Gem).
4. Name it something the whole desk will recognise: House Style — [Outlet Name].
5. Paste the style guide into the Instructions box. That is the large field asking what the Gem should do.
6. Save. Share it with the desk if your account allows sharing.

Step three: test it on real wire copy

We tested it the way the desk would actually use it — on agency copy that arrives already written and needs restyling.

1. Open the Associated Press of Pakistan website and pick any story off the wire.
2. Copy the text.
3. Open your new Gem and paste the story with a single line above it: “Turn this into a news story in our house style.”
4. Read the output against the original. Then read the MISSING list underneath it.

That MISSING list is the reason the Gem earns its place. Restyling copy is a convenience. Being told, every single time, that the story has no second source, or that a figure has arrived with no origin attached, or that the person being criticised has not been offered a right of reply — that is an editing standard running automatically on every piece of copy that crosses the desk, at three in the morning, when the person who would have caught it has gone home.

One warning that belongs on the wall next to it. Restyling wire copy never removes the obligation to credit the wire.The Gem is instructed not to strip that attribution, but instructions are not a guarantee. Check that the agency credit survived, every time. A house style that quietly erases where the story came from is not a style guide. It is a liability.

Tool four: turn a recurring page into an interactive dashboard, using Gemini Canvas

The first three tools help journalists produce the work. This one changes what the reader receives.

Every news website carries features that publish on a fixed schedule in a fixed shape. On this outlet’s Urdu site, the daily horoscope was one of them. It ran every day, it looked identical every day, and it was delivered the way such things have been delivered since print: a long block of text containing twelve entries, of which any given reader wants exactly one.

That is the flaw worth seeing. The reader is not looking for the horoscope. The reader is looking for their sign, and we make them scroll past eleven other people’s to reach it. The content is fine. The delivery is a leftover from a page that could not do anything else.

A web page can do something else. So we rebuilt it as an interactive dashboard, in one sitting, using Gemini Canvas — the mode of Gemini that writes and previews working web pages rather than describing them.

How we did it

1. Open the day’s feature on your own site and copy the content exactly as published.
2. Open Gemini in a new tab, start a fresh chat, and switch on Canvas.
3. Paste the master prompt below, with your content dropped into it, and send.
4. Look at the preview Canvas builds. Ask for changes in plain language — “make the cards bigger on mobile”, “move the date to the top” — until it matches what you want.

You are a front-end developer building a small interactive feature for a Pakistani news website.

Below is the content of a recurring daily feature from our site. Turn it into a single, self-contained, mobile-first interactive dashboard that we can publish.

CONTENT:

[paste the copied content here]

LANGUAGE: Urdu — the layout must run right-to-left, and the type must use a web font that renders Urdu correctly, loaded from Google Fonts. [If the feature is in English, replace this line with: LANGUAGE: English — left-to-right.] BUILD IT TO THESE RULES: 1. ONE FILE. All HTML, CSS and JavaScript in a single file, with no libraries or frameworks. The only outside request permitted is the web font. It must work if someone pastes it into a blank page. 2. MOBILE FIRST. Assume most readers arrive on a mid-range Android phone on a slow connection. Design for a 360-pixel-wide screen first and scale up from there. Tap targets at least 44 pixels. Keep the total weight as small as you can. No large images. 3. INTERACTIVE, NOT DECORATIVE. The reader must be able to do something — tap to open their own item, switch between items, reveal detail. Only one item’s detail is open at a time, so nobody scrolls past eleven entries to reach their own. 4. USE ONLY THE CONTENT I PASTED. Do not write, extend, embellish, translate or invent any of the text. If something is missing from what I pasted, leave the space empty and mark it clearly rather than filling it in yourself. 5. EASY TO UPDATE. Put every piece of daily text into one clearly labelled block at the very top of the file, so a desk editor can paste tomorrow’s content over yesterday’s without touching any code. Write a short comment above that block, in plain language, saying exactly what to change. 6. DATE STAMP. Show the date at the top, in our format, drawn from the block above so it updates with the content. 7. READABLE. Body text no smaller than 16 pixels. Strong contrast. Still works when the reader has set their phone’s font size large. Every interactive element reachable by keyboard and labelled for a screen reader. 8. NEUTRAL NEWS STYLING. Clean and restrained, so it can sit inside an existing news page without clashing. No animation beyond a simple open and close. One accent colour, defined once at the top so it can be changed in a second. Then, separately from the code and in plain non-technical language, tell me: – exactly how to put this on a WordPress website, step by step, in the order I should click – what to do if the layout looks broken inside our page template – how a desk editor updates it tomorrow, in three steps or fewer

The last part of that prompt matters as much as the code. I did not want the desk holding a file it could not deploy, so the prompt makes Gemini teach the embedding as well as write the dashboard. The person who runs the feature should be able to publish it and update it without filing a ticket with anyone.

How to find the next four

Once the horoscope was rebuilt, we went looking for the rest, and identified at least five recurring features on that site that could be converted the same way. Three questions decide it:

• Does it publish on a fixed schedule? Daily, weekly, or every match day.
• Is its shape identical every time? The same categories, in the same order, with only the values changing.
• Does the reader want one part of it rather than all of it? Their sign. Their city. Their currency. Their team.

Three yes answers and the feature is a dashboard being delivered as an article. Anything that fails the third question should stay as it is — interactivity added to content people genuinely read end to end is friction, not innovation.

The intention here is straightforward: a reader who can tap straight to the part that concerns them has a better experience and stays longer than one made to scroll through eleven entries that do not. I will not put a number on that improvement, because we did not measure it, and a number I have not measured is a number I will not publish. What I can say is that the raw material was already there. It was being published every day, in the least useful shape available.

What rules should sit above every AI tool in a newsroom?

No prompt replaces judgement. Five rules travel with every use, and I say them out loud in every room.

The number in the headline gets checked against the original source. Not against the AI’s answer — the document itself. It takes ten seconds and it is the entire difference between a story and a correction.

Never paste confidential material into any consumer AI tool. Nothing unpublished, nothing that could identify a source, no private personal data, nothing under legal review. If you would not put it on the photocopier in the lobby, do not put it in a chatbot.

The right of reply is not optional. If a story criticises anyone, the prompt lists them as a call. Make the call.

Keep it clear of the political line. Technology, education, showbiz, business, sports and health are everyday beats. If a scan surfaces something political, security-related, or a sensitive controversy, that is an editorial decision for your desk. Treat it exactly as you would any wire copy on a sensitive topic, not as something to publish as it came out of the tool.

Wire copy keeps its credit. Restyling an agency story through your own house-style assistant never removes the obligation to say where it came from. Check the credit survived, every time.

What happens next: the master class is now inevitable

I went into that room to test a training. I came out of it with a mandate I did not design for.

Ninety-six percent of a working newsroom asked for more. The least confident asked most unanimously of all. Every senior journalist in the room said yes. And the pattern in the data — decisive at the top of the experience curve, hedging in the middle of it — tells me exactly what a longer programme has to do: give the newer journalists the map they do not yet have, and give the experienced ones the specific workflow proof that turns a considered “no” into a considered “yes.”

A single workshop cannot do that. A written prompt system, like the one above, gets a desk moving by Monday, but it does not build judgement, and judgement is the actual skill. So I am building the longer programme.

I am now convinced that a master class is not optional, and I will be announcing the launch of a Master Class in AI shortly, alongside programmes in a small number of adjacent domains. It is being designed around what this data showed rather than around what I assumed before I ran the survey: structured for the hesitant rather than the enthusiasts, workflow-first rather than tool-first, and built so a journalist finishes it able to interrogate any system that has not been invented yet.

It is being built for three audiences. Students and early-career journalists, who need the discipline installed before the habits set. Journalism institutes and universities, who need a module that teaches judgement rather than a software demonstration that expires with the next product update. And NGOs and mission-driven communications teams, drowning in reports, policy documents and field data, who need the same interrogation method pointed at their material instead of the news wire.

If you want to be told first when enrolment opens, or you want this delivered inside your newsroom, institute or organisation before then, the contact page on this site is where that starts.

Frequently asked questions

---

Written by Hissan Gul, a strategic communications and audience intelligence practitioner with more than fifteen years across newsrooms, the development sector, e-commerce, paid media and training. He is a Fellow of the International Center for Journalists (ICFJ) and a Google-certified Master Trainer for the Google News Initiative. Published September 2026.