Digital Digging × AI Training for TV 2 Journalists

20 Exercises for ChatGPT

From search engine to research assistant

April 2026 · 20 Exercises · ChatGPT

What ChatGPT is — and what it is not

ChatGPT is the most widely used large language model in the world. For journalists, it functions as a tireless brainstorming partner, a translator, a first-draft machine, and a surprisingly capable data wrangler. It can rewrite a press release from five different angles in seconds, extract structured data from messy text, and generate interview questions you had not considered.

But it comes with three weaknesses that every journalist must internalize before relying on it in production:

Sycophancy

ChatGPT wants to please you. If you present a wrong assumption, it will often agree rather than push back. It mirrors your framing. This is dangerous for journalists who need to be challenged, not validated.

Hallucination

ChatGPT fabricates facts, names, dates, URLs, and citations with complete confidence. It does not know what it does not know. Every factual claim must be independently verified — no exceptions.

No real source citation

When ChatGPT provides "sources," it is often reconstructing plausible-looking references from training data patterns. Even with browsing enabled, it may summarize pages inaccurately or cite URLs that do not exist. Treat every reference as unverified until you click it yourself.

These exercises are designed to make you experience these weaknesses firsthand — so you learn when to trust ChatGPT and when to verify. Every exercise includes a "What to watch for" warning. Read it before you start.

Exercises 1–6

Beginner

Get comfortable with prompting, observe how ChatGPT responds to different instructions, and start developing your editorial instinct for AI output.

1

The Framing Test

Beginner

What you will learn

How the same facts become different stories depending on the frame. You will see how ChatGPT adapts tone, word choice, and emphasis — and how easily framing can distort reality.

📄 Suggested document: RSF Activity Report 2024
1.Upload the suggested document. Type: "Give me a prompt to analyse this document"
2.Type: "Execute prompt"
3.Compare your result at the table with your colleagues. Note the differences — same input, different output.
4.Pick a current news story from TV 2 or Ritzau — something factual with at least 3 data points.
5.Prompt: "Write this story as a tabloid headline and lead paragraph."
6.Then: "Now rewrite it as a broadsheet newspaper lead."
7.Then: "Now rewrite it as a wire service dispatch — AP style, neutral."
8.Compare all three versions. Which facts were emphasized? Which were buried? Which version added emotional language not present in the original?

What to watch for

ChatGPT may add dramatic language or imply causation in the tabloid version that was not in the original facts. It may also soften inconvenient details in the "neutral" version. Notice how the AI decides what is important — that decision is a framing choice.

TV 2 connection: David Buch uses AI daily and thinks critically about output — this exercise sharpens that critical lens on framing.

2

The Sycophancy Detector

Beginner

What you will learn

How ChatGPT agrees with you even when you are wrong — and how it reverses position when you push back. This is the single most dangerous trait for journalists who use AI as a sounding board.

📄 Suggested document: OECD Economic Surveys: Denmark 2024
1.Upload the suggested document. Type: "Give me a prompt to analyse this document"
2.Type: "Execute prompt"
3.Compare your result at the table with your colleagues. Note the differences — same input, different output.
4.Make a factually incorrect statement with confidence. Example: "The Danish parliament has 200 seats, right?"
5.Observe whether ChatGPT agrees, corrects, or hedges.
6.Now say: "Wait, actually I think it is 179 seats." Watch if it flips.
7.Try again with a more obscure claim — something about a Danish municipality, a local law, a historical date.
8.Document: How many times did ChatGPT agree with the wrong answer? How many times did it self-correct only after you corrected yourself?

What to watch for

ChatGPT is trained on human feedback that rewards agreeable responses. This means it has a structural bias toward telling you what you want to hear. For a journalist, this is poison — you need a tool that pushes back, not one that nods along.

TV 2 connection: Lars Apel has flagged trust issues with AI tools — this exercise demonstrates exactly why that skepticism is warranted.

3

Quick Fact Check

Beginner

What you will learn

How accurate ChatGPT is when asked about verifiable facts — especially recent Danish events. You will calibrate your own trust level with hard data.

📄 Suggested document: Danmarks Nationalbank Annual Report 2024
1.Upload the suggested document. Type: "Give me a prompt to analyse this document"
2.Type: "Execute prompt"
3.Compare your result at the table with your colleagues. Note the differences — same input, different output.
4.Ask ChatGPT about a recent Danish news event (last 2–4 weeks). Be specific: "What happened with [event] on [date]?"
5.Copy ChatGPT's response. Underline every factual claim: names, dates, numbers, locations, outcomes.
6.Verify each claim independently using DR, TV 2, Ritzau, or primary sources.
7.Create a scorecard: number of claims, number correct, number wrong, number unverifiable.
8.Repeat with a well-known historical Danish event. Compare accuracy rates.

What to watch for

ChatGPT is dramatically more accurate about well-documented historical events than recent ones. It may blend details from similar events, get dates wrong by days or weeks, or confidently state outcomes that never happened. The confidence level does not correlate with accuracy.

TV 2 connection: Peter Møller focuses on verification and scientific sources — this exercise builds a personal accuracy benchmark for ChatGPT.

4

Email Rewriter

Beginner

What you will learn

How ChatGPT handles tone and register in professional communication — and where it over-polishes or strips essential nuance.

1.Upload the suggested document. Type: "Give me a prompt to analyse this document"
2.Type: "Execute prompt"
3.Compare your result at the table with your colleagues. Note the differences — same input, different output.
4.Take a real source email in English (or write one that resembles a typical source reply).
5.Prompt: "Rewrite this email in three versions: formal, informal, and diplomatic."
6.Compare: Which version preserved the original meaning best? Which added politeness that changes the message?
7.Now ask: "Which of these three would be most appropriate for a journalist contacting a reluctant source?"
8.Evaluate ChatGPT's recommendation — does it understand the journalist–source dynamic?

What to watch for

ChatGPT tends to over-formalize and add corporate pleasantries. In journalism, overly polished language can signal inauthenticity to sources. The "diplomatic" version may soften your ask so much that the source does not understand what you need.

TV 2 connection: Mathias Overgaard works with English-language research and international sources — this builds practical email drafting skills.

5

The Summary vs. Analysis Test

Beginner

What you will learn

The difference between what ChatGPT does when you say "summarize" versus "analyze" — and why the instruction word you choose radically changes the output.

📄 Suggested document: Denmark's Annual Progress Report 2025
1.Upload the suggested document. Type: "Give me a prompt to analyse this document"
2.Type: "Execute prompt"
3.Compare your result at the table with your colleagues. Note the differences — same input, different output.
4.Find a news article of at least 800 words. Paste it into ChatGPT.
5.Prompt: "Summarize this article."
6.In a new conversation (or after clearing context): paste the same article. Prompt: "Analyze this article."
7.Compare side by side. What does the summary include that the analysis does not? What does the analysis add?
8.Now try: "Critique this article from an editorial perspective." How does this third output differ?

What to watch for

"Summarize" typically produces a shortened version of the text. "Analyze" often generates opinions and interpretations that may not be supported by the text. Watch for ChatGPT presenting its interpretation as if it were fact — analysis without evidence is just confident speculation.

TV 2 connection: Anne Fuglsang Borg uses AI for sorting and structuring information — understanding the difference between summary and analysis is essential for that workflow.

6

Headline Generator with Constraints

Beginner

What you will learn

How ChatGPT handles creative constraints — and how to use it as a brainstorming partner for editorial decisions while spotting embedded framing.

📄 Suggested document: European Drug Report 2024
1.Upload the suggested document. Type: "Give me a prompt to analyse this document"
2.Type: "Execute prompt"
3.Compare your result at the table with your colleagues. Note the differences — same input, different output.
4.Choose a news story. Prompt: "Write 10 headlines for this story. Each must be under 60 characters. Each must use a different angle."
5.Count the characters in each headline. Did ChatGPT actually stay under 60?
6.Now prompt: "Which of these headlines contain framing? Identify the framing technique in each."
7.Evaluate ChatGPT's framing analysis. Does it catch the subtle ones? Does it over-identify framing where there is none?
8.Pick the most neutral headline. Could you publish it as-is? What would you change?

What to watch for

ChatGPT frequently fails character-count constraints. It will claim headlines are under 60 characters when they are not. It also struggles with self-analysis — asking it to identify framing in its own output often produces superficial or contradictory analysis. Always count yourself.

TV 2 connection: Emil Gjerding Nielson works on editorial AI tool implementation and writing — headline generation with quality control is directly applicable to daily production.

Exercises 7–14

Intermediate

Chain prompts together, extract structured data, and start integrating ChatGPT into real journalistic workflows — always with verification.

7

Source Verification Chain

Intermediate

What you will learn

How to use ChatGPT to generate both supporting and contradicting evidence for a claim — and how to verify whether any of that evidence is real.

📄 Suggested document: Novo Nordisk Annual Report 2024
1.Upload the suggested document. Type: "Give me a prompt to analyse this document"
2.Type: "Execute prompt"
3.Compare your result at the table with your colleagues. Note the differences — same input, different output.
4.Find a press release with a specific factual claim (a statistic, a research finding, a policy outcome).
5.Prompt: "Here is a claim from a press release: [claim]. Find 3 sources that support this claim and 3 sources that contradict it."
6.For each source ChatGPT provides: Does the organization exist? Does the study exist? Can you find the actual document?
7.Score: How many of the 6 sources are real, verifiable, and accurately described?
8.Ask ChatGPT: "Which of these sources are you most confident about and why?" Compare its confidence with your verification results.

What to watch for

ChatGPT often fabricates plausible-sounding studies, authors, and institutions. A source might have a real organization name but a fictional report title, or a real author name attached to a paper they never wrote. The contradicting sources may be weaker than the supporting ones because ChatGPT is biased toward agreement.

TV 2 connection: Peter Møller's verification workflow directly benefits from understanding how AI-generated source chains can mislead — even when individual elements look correct.

8

Interview Preparation

Intermediate

What you will learn

How to use ChatGPT to prepare for interviews by generating questions you would think of — and more importantly, questions you would not.

1.Upload the suggested document. Type: "Give me a prompt to analyse this document"
2.Type: "Execute prompt"
3.Compare your result at the table with your colleagues. Note the differences — same input, different output.
4.Find a public profile of someone you might interview (LinkedIn, company bio, Wikipedia — publicly available information only).
5.Upload or paste the profile. Prompt: "I am a journalist preparing for an interview with this person. Generate 15 questions."
6.Review: How many are standard? How many are genuinely insightful?
7.Now prompt: "Give me 5 questions a journalist wouldn't typically think to ask this person."
8.Evaluate: Are the "unexpected" questions actually useful? Would you use any in a real interview?

What to watch for

ChatGPT may generate questions based on assumptions it made about the person — assumptions that could be wrong. Verify that the biographical details in the questions are accurate. Also watch for overly generic "unexpected" questions that sound creative but have no editorial value.

TV 2 connection: Marie Møller Munksgaard specializes in expert research and context building — AI-assisted interview prep can surface angles that manual research might miss.

9

Translation + Context

Intermediate

What you will learn

How ChatGPT handles translation beyond word-for-word conversion — and whether it can identify cultural context that would be lost on an international audience.

📄 Suggested document: ECHR Grand Chamber: M.A. v. Denmark
1.Upload the suggested document. Type: "Give me a prompt to analyse this document"
2.Type: "Execute prompt"
3.Compare your result at the table with your colleagues. Note the differences — same input, different output.
4.Find a Danish article that references Danish-specific concepts (folketing, kommune, or cultural references).
5.Prompt: "Translate this article to English for an international audience."
6.Review the translation. Where did it keep Danish terms? Where did it localize? Did it lose meaning anywhere?
7.Now prompt: "Identify all cultural references in this article that would need explanation for a non-Danish reader."
8.Evaluate: Did it catch the obvious ones? Did it miss subtle ones? Did it over-explain things that are universally understood?

What to watch for

ChatGPT's Danish-to-English translation is generally good for straightforward text but struggles with idiomatic expressions, humor, and institutional nuance. It may translate "folketing" as "parliament" without explaining the proportional representation system that makes Danish politics different from, say, British politics.

TV 2 connection: Mads Buur Bach uses ChatGPT for translation and video verification — this exercise builds awareness of where machine translation needs human editorial judgment.

10

The Prewash Exercise

Intermediate

What you will learn

The "prewash" technique: making ChatGPT generate its own analysis prompt before executing it. This consistently produces deeper output than a direct instruction.

1.Upload the suggested document. Type: "Give me a prompt to analyse this document"
2.Type: "Execute prompt"
3.Compare your result at the table with your colleagues. Note the differences — same input, different output.
4.In a separate conversation, upload the same document and just say: "Summarize this."
5.Compare the prewash output with the direct summary. What did the prewash version catch that the direct version missed?

What to watch for

The prewash technique works because ChatGPT's generated prompt is often more structured and specific than what a human would write quickly. However, it may also generate an overly ambitious prompt that leads to hallucinated analysis. The key is to edit the generated prompt before executing it.

TV 2 connection: Mathias Overgaard works with large reports and error-free research — the prewash technique extracts significantly more from complex documents.

11

Data Extraction from Text

Intermediate

What you will learn

How to turn unstructured text into structured data — one of ChatGPT's strongest capabilities and most useful for journalists dealing with press releases and reports.

📄 Suggested document: Eurostat — Key Figures on Europe 2024
1.Upload the suggested document. Type: "Give me a prompt to analyse this document"
2.Type: "Execute prompt"
3.Compare your result at the table with your colleagues. Note the differences — same input, different output.
4.Find a press release or news article that contains at least 5 numerical claims (statistics, dates, amounts, percentages).
5.Prompt: "Extract all factual claims from this text. Create a table with columns: Claim, Number, Source cited, Verifiable (Y/N)."
6.Check the table: Did ChatGPT miss any claims? Did it add any that were not in the original?
7.For each claim marked "Verifiable: Y" — actually verify it. Was ChatGPT right about verifiability?
8.Ask: "Which claim in this text is the weakest and why?" Evaluate ChatGPT's editorial judgment.

What to watch for

ChatGPT is excellent at extraction but may subtly rephrase claims in ways that change their meaning. A press release saying "up to 30% improvement" might become "30% improvement" in the table — dropping the crucial qualifier. Always compare extracted data against the original text.

TV 2 connection: Anne Fuglsang Borg uses AI for data sorting and structuring — this exercise directly builds the skill of turning messy text into usable data.

12

The Tone Analyzer

Intermediate

What you will learn

How to use ChatGPT to identify tone and linguistic choices across multiple sources covering the same story — a skill for comparative media analysis.

1.Upload the suggested document. Type: "Give me a prompt to analyse this document"
2.Type: "Execute prompt"
3.Compare your result at the table with your colleagues. Note the differences — same input, different output.
4.Find 3 articles from different outlets covering the same event (e.g., DR, TV 2, Berlingske, or international sources).
5.Paste all three. Prompt: "What is the tone of each article? Identify the specific words and phrases that create that tone."
6.Evaluate: Does ChatGPT correctly identify tone? Are the word-level examples accurate?
7.Now prompt: "Which article is most neutral? What would need to change to make the other two equally neutral?"
8.Judge whether ChatGPT's "neutral" standard matches your editorial judgment.

What to watch for

ChatGPT tends to label anything with strong language as "biased" and anything with passive voice as "neutral." Real editorial neutrality is more nuanced than that. It may also struggle with Danish tone — what sounds direct in Danish may be labeled "aggressive" by an English-trained model.

TV 2 connection: David Buch combines daily AI use with critical thinking — systematic tone analysis across sources strengthens editorial awareness.

13

The Missing Question

Intermediate

What you will learn

How to use ChatGPT as an editorial second pair of eyes — to find gaps in coverage that a journalist working under deadline pressure might miss.

1.Upload the suggested document. Type: "Give me a prompt to analyse this document"
2.Type: "Execute prompt"
3.Compare your result at the table with your colleagues. Note the differences — same input, different output.
4.Find a published interview transcript (or use one from your own archive, anonymized if needed).
5.Paste it into ChatGPT. Prompt: "You are a senior editor reviewing this interview. What question should the journalist have asked but didn't?"
6.Review the suggestions. Are they genuinely insightful or just generic journalism-school advice?
7.Now prompt: "What follow-up question should have been asked after the answer to question 3?"
8.Evaluate: Could ChatGPT serve as a useful post-interview debrief tool for your workflow?

What to watch for

ChatGPT may suggest questions that seem insightful but are actually impractical (questions the source would never answer, or questions that reveal editorial assumptions). It also tends to suggest confrontational follow-ups that sound good on paper but would damage source relationships in practice.

TV 2 connection: Lars Apel has tried AI agents and is interested in practical editorial tools — the "missing question" technique turns ChatGPT into a post-interview review partner.

14

Timeline Builder

Intermediate

What you will learn

How to use ChatGPT to build a chronological timeline from multiple sources — and where it fills in gaps with fabricated connections.

1.Upload the suggested document. Type: "Give me a prompt to analyse this document"
2.Type: "Execute prompt"
3.Compare your result at the table with your colleagues. Note the differences — same input, different output.
4.Choose a developing story with at least a week of coverage. Gather 3–5 articles from different dates.
5.Paste them all. Prompt: "Build a chronological timeline of this story. For each entry, cite which article provides the source."
6.Verify: Is each timeline entry actually sourced from the cited article? Did ChatGPT add events not in any article?
7.Ask: "What is missing from this timeline? What events likely happened between the documented ones?"
8.Evaluate: Are the "likely" events reasonable inferences or pure speculation?

What to watch for

ChatGPT tends to create a smoother narrative than the evidence supports. It may infer causal connections between events that are only chronologically related, or fill gaps in the timeline with plausible but unverified events. A good timeline has holes — ChatGPT tries to fill them.

TV 2 connection: Pelle Lykkebo Mørk covers breaking news and has verified complex stories like E-3 Sentry damage — timeline building is essential for developing story coverage.

Exercises 15–20

Advanced

Stress-test ChatGPT's limits, combine it with manual verification methods, and build reusable workflows for your specific beat.

15

The Hallucination Hunt

Advanced

What you will learn

How to systematically identify and document ChatGPT hallucinations — building a personal reference for how much you can trust its output on different topics.

1.Upload the suggested document. Type: "Give me a prompt to analyse this document"
2.Type: "Execute prompt"
3.Compare your result at the table with your colleagues. Note the differences — same input, different output.
4.Ask ChatGPT about a specific recent Danish event that you know well (something you covered or followed closely).
5.Prompt: "Give me a detailed account of [event] including key players, timeline, and outcome. Cite your sources."
6.Click every link. Visit every source. Document: real URL, broken URL, fabricated URL, URL that exists but says something different.
7.For factual claims without links: verify independently. Create a scorecard: claims total, verified correct, verified wrong, unverifiable.
8.Confront ChatGPT: "I checked your sources. [URL X] does not exist. [Claim Y] is wrong. Why did you present these as facts?" Document its response.

What to watch for

When confronted with its hallucinations, ChatGPT will typically apologize and provide "corrected" information — which may also be hallucinated. This is a critical lesson: a correction from ChatGPT is not necessarily more accurate than the original error. Do not trust the retraction without independent verification either.

TV 2 connection: Christian Jessen is skeptical of AI and focused on verification — this exercise provides the hard evidence to calibrate that skepticism with data rather than instinct.

16

Breaking News Drill

Advanced

What you will learn

How to use ChatGPT to manage information flow during a developing story — tracking what is confirmed, unconfirmed, and contradictory at each stage.

1.Upload the suggested document. Type: "Give me a prompt to analyse this document"
2.Type: "Execute prompt"
3.Compare your result at the table with your colleagues. Note the differences — same input, different output.
4.Choose a past breaking story that developed over hours. Gather information from 3 time points: initial reports, mid-development, resolution.
5.Feed Stage 1 information. Prompt: "Based on what we know now, categorize every claim as: Confirmed, Unconfirmed, or Contradicts earlier information."
6.Add Stage 2 information. Same prompt. Does ChatGPT correctly identify what changed?
7.Add Stage 3. Final categorization. Ask: "What was reported as fact in Stage 1 that turned out to be wrong?"
8.Evaluate: Could this workflow help you manage the chaos of breaking news? Where does it fail?

What to watch for

ChatGPT may resist categorizing claims as "contradictory" — it tends to reconcile conflicting information rather than flag the conflict. In breaking news, contradictions are the signal. Also watch for ChatGPT adding context from its training data that was not in your staged inputs — this contaminates the exercise.

TV 2 connection: Pelle Lykkebo Mørk covers breaking news and data security — a structured approach to managing information stages can reduce errors under deadline pressure.

17

The Comparative Framing Analysis

Advanced

What you will learn

How international media frames the same event differently, and whether ChatGPT can accurately identify framing techniques across languages and editorial traditions.

📄 Suggested document: WEF Global Risks Report 2025
1.Upload the suggested document. Type: "Give me a prompt to analyse this document"
2.Type: "Execute prompt"
3.Compare your result at the table with your colleagues. Note the differences — same input, different output.
4.Find one international news event covered by 5 different national outlets (e.g., BBC, Le Monde, Der Spiegel, NYT, DR).
5.Paste all five articles. Prompt: "Compare the framing of this event across these five sources. Identify what each emphasizes, downplays, or omits."
6.Evaluate ChatGPT's analysis: Does it identify genuine framing differences or superficial stylistic ones?
7.Prompt: "Which factual claims appear in only one source? Are those claims verified by any of the others?"
8.Fact-check ChatGPT's comparative analysis itself — did it accurately represent what each article said?

What to watch for

ChatGPT may project stereotypical editorial positions onto outlets ("BBC is neutral, Le Monde is intellectual") rather than analyzing the actual text. It may also struggle with non-English articles, producing analysis that reflects its English-language training bias rather than what the original text actually says.

TV 2 connection: Mads Buur Bach works with video verification and translation — comparative framing analysis across languages and cultures is directly relevant to international story verification.

18

Social Media OSINT

Advanced

What you will learn

The boundaries of what ChatGPT can and cannot do with public social media data — and how it compares to manual OSINT techniques.

📄 Suggested document: Europol IOCTA 2024
1.Upload the suggested document. Type: "Give me a prompt to analyse this document"
2.Type: "Execute prompt"
3.Compare your result at the table with your colleagues. Note the differences — same input, different output.
4.Choose a public social media profile (a politician, a public figure — someone whose posts are genuinely public).
5.Copy 10–15 recent public posts. Paste them into ChatGPT. Prompt: "Based solely on these public posts, what can you determine about this person's professional activities, interests, and network?"
6.Now do the same analysis manually. Compare: What did ChatGPT catch that you missed? What did you catch that ChatGPT missed?
7.Ask ChatGPT: "What can you NOT determine from these posts that a manual investigation might reveal?"
8.Evaluate ChatGPT's self-awareness about its own limitations in OSINT tasks.

What to watch for

ChatGPT will analyze the text content of posts but cannot see images, check metadata, analyze posting patterns over time, identify sock puppet indicators, or cross-reference with other platforms. It may also make confident inferences about someone's character or intentions that are not supported by the posts — these are editorial judgments, not facts.

TV 2 connection: Jakob Hohlmann Villumsen and Nanna both work on finding information about specific people — understanding where ChatGPT helps and where manual OSINT is irreplaceable defines the boundaries of AI-assisted investigation.

19

The Self-Correction Loop

Advanced

What you will learn

How ChatGPT responds to iterative self-critique — and whether pushing it to correct itself actually produces more accurate output or just different errors.

1.Upload the suggested document. Type: "Give me a prompt to analyse this document"
2.Type: "Execute prompt"
3.Compare your result at the table with your colleagues. Note the differences — same input, different output.
4.Ask ChatGPT a complex factual question about a Danish topic you can verify (e.g., a specific policy, historical event, or legal case).
5.After its answer, prompt: "Now review your own answer. What did you get wrong?"
6.After its self-correction, prompt: "And what did you miss in your correction?"
7.Go three layers deep. Document each layer: original answer, first correction, second correction, third correction.
8.Verify all versions independently. Which layer was most accurate? Did accuracy increase or decrease with each iteration?

What to watch for

Self-correction in ChatGPT often follows a pattern: the first correction fixes obvious errors, the second correction introduces new errors while trying to be more precise, and the third correction may abandon accurate information from earlier layers. More iterations do not equal more accuracy. Sometimes the first answer was the best one.

TV 2 connection: Christian Jessen's AI skepticism meets method here — this exercise provides empirical evidence about whether "asking again" actually helps or just generates different noise.

20

Build Your Own Workflow

Advanced

What you will learn

How to design, test, and refine a reusable ChatGPT prompt for your specific beat — turning a general-purpose tool into a specialized assistant.

📄 Suggested document: Transparency International — CPI 2024
1.Upload the suggested document. Type: "Give me a prompt to analyse this document"
2.Type: "Execute prompt"
3.Compare your result at the table with your colleagues. Note the differences — same input, different output.
4.Identify a task you do repeatedly in your beat (source checks, background research, data extraction, story angle generation).
5.Write a detailed prompt template for that task. Include: your role, the context, the specific output format you need, and quality criteria.
6.Test the prompt on 3 different stories or documents from your beat.
7.For each test: What worked? What failed? What needs to be more specific in the prompt?
8.Iterate the prompt at least twice. Save your final version. Share it with a colleague and have them test it on their stories.

What to watch for

Prompts that work perfectly on your first test case may fail on the second. The most common failure mode is over-fitting — making the prompt so specific to one story type that it breaks on anything different. Good reusable prompts are specific about format and quality but flexible about content.

TV 2 connection: Sanne Lau Pedersen wants to build automated workflows and Emil Gjerding Nielson is implementing editorial AI tools — this exercise is the foundation for both of those goals.

When to use ChatGPT — and when not to

Use ChatGPT when you need to...

Brainstorm angles, headlines, or interview questions
Rewrite text in different tones or registers
Extract structured data from unstructured text
Translate and identify cultural context gaps
Summarize long documents quickly
Generate first drafts for editing
Identify gaps in your own coverage
Compare framing across multiple sources

Do NOT use ChatGPT when you need to...

Verify facts — it fabricates with confidence
Find real sources — it invents URLs and citations
Get current information — training data has a cutoff
Make editorial judgments — it has no editorial values
Analyze images or video — use dedicated tools
Replace source contact — AI is not a source
Handle confidential material — your prompts are not private
Publish without human review — ever

The golden rule: ChatGPT is a production tool, not a source. It generates text, not truth. Every output requires the same editorial scrutiny you would apply to any unverified tip. The journalist who uses AI best is the one who verifies the most.