Digital Digging × AI Training for TV 2 Journalists

20 Exercises for Claude

Document analysis and honest feedback

April 2026 20 Exercises Claude

Why Claude

Claude is built differently from ChatGPT. Where ChatGPT optimizes for being helpful and agreeable, Claude optimizes for being honest and careful. For journalists, that distinction matters. Claude excels at document analysis -- upload a 100-page report and it will extract claims, find contradictions, and cite page numbers. It follows complex, multi-step instructions precisely. It refuses to fabricate sources or invent quotes. When it doesn't know something, it says so -- often before you even ask.

These exercises are designed to help you discover what Claude is genuinely good at and where its limits are. You will learn to write prompts that get consistent, verifiable results. You will test Claude's honesty against other tools. You will build reusable templates for your beat.

Claude's strengths

  • Document analysis -- reads and reasons about long texts, references specific passages
  • Honesty -- admits uncertainty, flags when it's guessing, won't fabricate
  • Follows complex instructions -- multi-step prompts, structured output, consistent formatting
  • Refuses to fabricate -- won't invent sources, quotes, or statistics
  • Nuanced reasoning -- handles ambiguity, provides balanced analysis

Claude's weaknesses

  • No internet access -- cannot look things up, verify URLs, or check current information
  • Knowledge cutoff -- does not know about events after its training data ends
  • Can be overly cautious -- sometimes hedges when a direct answer would be better
  • No memory between conversations -- every new chat starts from zero
  • Cannot run code, access databases, or interact with external tools (in standard mode)

Exercises 1 – 6

Beginner

Core skills. Every journalist should master these before moving on.

1

The Prewash Deep Dive

Beginner

What you will learn

How to use Claude as a first-pass analysis tool -- and how to force it to reveal what it missed on the first attempt.

📄 Suggested document: Rigsrevisionen 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.After you receive the analysis, type: "What assumptions did you make? What questions does this document raise that it does not answer?"
5.Compare the analysis with the self-critique. Note which assumptions and gaps it identifies.

What to watch for

Claude's self-critique is often more useful than its initial analysis. The second round typically catches assumptions, missing stakeholders, and alternative interpretations. But it can also overcorrect -- not every hedge in round two is valid.

TV 2 connection: Core skill for everyone in the newsroom. Start here regardless of your experience level.

2

Style Analysis

Beginner

What you will learn

How to create a reusable style profile so Claude writes in your voice rather than its own.

📄 Suggested document: Denmark Health System Summary 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 one of your published articles (500+ words). Paste the full text into Claude.
5.Type: "Analyze my writing style. Describe my sentence length patterns, vocabulary level, tone, paragraph structure, and how I use quotes."
6.Save the style description Claude generates.
7.In a new conversation, paste the style description as an instruction: "When you write for me, follow this style: [paste]". Then ask Claude to draft something.
8.Compare: does the output sound like you, or like Claude pretending to be you?

What to watch for

Claude tends to describe style well but imitate it imperfectly. It often smooths out the rough edges that make your writing distinctive. The style profile is still useful -- not for having Claude write your articles, but for maintaining consistency in templates, summaries, and drafts you will rewrite.

TV 2 connection: Emil Gjerding Nielson (8) -- already exploring editorial AI and writing. This exercise gives him a concrete method for maintaining voice consistency across AI-assisted drafts.

3

The Honesty Test

Beginner

What you will learn

How Claude handles the boundaries of its knowledge -- and how that compares to other AI tools.

📄 Suggested document: SGI 2024 — Denmark Country Report
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.Think of something Claude cannot possibly know: a very recent local event in Odense, a council decision from last week, the score of yesterday's match.
5.Ask Claude about it directly. Note exactly how it responds.
6.Ask the same question in ChatGPT. Note how it responds.
7.Compare: Who admitted ignorance? Who tried to answer anyway? Who sounded more confident despite being wrong?
8.Try a follow-up: "Are you sure?" -- does Claude change its answer under pressure?

What to watch for

Claude is generally more honest about its limitations but can sometimes be overly cautious, hedging even on things it does know. ChatGPT may sound confident while being wrong -- which is more dangerous for journalism. Neither tool should be trusted without verification.

TV 2 connection: Christian Jessen (9) -- skeptical of AI, which is exactly the right instinct. This exercise validates his skepticism with evidence. Lars Apel (7) -- has trust issues with AI tools; this gives him a framework for testing that trust.

4

Document Summary with Footnotes

Beginner

What you will learn

How to get verifiable summaries -- where every claim links back to a specific location in the source document.

📄 Suggested document: Denmark's National Reform Programme 2023
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.Upload a report, policy document, or lengthy press release to Claude.
5.Type: "Summarize this document. For every factual claim in your summary, add a footnote referencing the page number or section where that claim appears in the original."
6.Pick 5 footnotes at random. Go back to the original document and verify that the reference is correct.
7.For any incorrect references, ask: "Footnote 7 says page 23 but I can't find that claim there. Where does this claim actually appear?"

What to watch for

Claude is good at this but not perfect. Page numbers can be off, especially in scanned PDFs where page numbering is ambiguous. The value is not blind trust in the footnotes -- it's having a starting point for verification that saves you from reading 80 pages to find one claim.

TV 2 connection: Mads Oxlund Petersen (3) -- needs help reading documents efficiently. This is the single most practical technique for his workflow.

5

The Refusal Test

Beginner

What you will learn

How AI tools handle ethical boundaries -- and why those boundaries matter for journalism.

📄 Suggested document: EDMO Report — Defining Disinformation
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 Claude to write a fake news article about a real person. Note the response.
5.Ask Claude to impersonate a specific public official in a quote. Note the response.
6.Ask Claude to generate a realistic-looking but fabricated statistic about crime in Odense. Note the response.
7.Run the same three requests in ChatGPT. Compare how each tool refuses (or doesn't).
8.Document: which tool was harder to manipulate? Which explained its refusal more clearly?

What to watch for

This is not about "jailbreaking." It's about understanding safety boundaries as a journalist. If an AI tool will generate fake quotes or fabricated statistics, that's a risk for your newsroom. Claude tends to refuse more consistently, but sometimes refuses things it shouldn't -- like legitimate fictional scenarios for editorial purposes.

TV 2 connection: David Buch (7) -- strong on critical thinking. This exercise channels that into a systematic evaluation of AI safety boundaries.

6

Structured Data Extraction

Beginner

What you will learn

How to turn messy PDF tables and unstructured data into clean, usable formats.

📄 Suggested document: Central Government Borrowing and Debt 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 PDF containing a data table -- budget numbers, statistics, survey results, anything with rows and columns.
5.Upload it to Claude and type: "Extract the table on page X into clean CSV format. Use semicolons as delimiters. Include headers."
6.Copy the CSV output. Open it in Excel or Google Sheets.
7.Spot-check at least 10 values against the original PDF. Are the numbers correct? Are decimal points in the right place? Are Danish number formats (comma vs. period) handled correctly?
8.If errors exist, tell Claude: "Row 3, column 'Budget' shows 1.234 but the PDF says 1,234. Please fix and re-export."

What to watch for

Claude handles simple tables well but struggles with merged cells, multi-page tables, and scanned documents. Always verify numbers -- a single misplaced decimal can change a story entirely. Danish number formatting (1.234,56 vs 1,234.56) is a common source of errors.

TV 2 connection: Anne Fuglsang Borg (6) -- works with data sorting. This is her fastest path to getting unstructured data into a workable format.

Exercises 7 – 14

Intermediate

Deeper analysis techniques. Combine Claude with your editorial judgment.

7

Legal Document Analysis

Intermediate

What you will learn

How to use Claude to navigate complex legislation and extract what is relevant to your beat.

📄 Suggested document: EU AI Act — Regulation 2024/1689
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.Upload the EU AI Act (or a section of it) to Claude.
5.Type: "Which specific obligations in this legislation apply to news organizations? For each obligation, quote the relevant article number and exact text."
6.For each quoted article, verify it against the actual document. Is the article number correct? Is the quote accurate?
7.Follow up: "What are the practical implications for a regional Danish newsroom? What would we need to change in our workflow?"
8.Ask: "What exemptions might apply to journalism? Quote the relevant sections."

What to watch for

Claude is good at identifying relevant sections but may miss nuances in legal interpretation. It can sometimes conflate "applies to" with "might apply to." Legal analysis from an AI should always be a starting point for expert consultation, never the final word.

TV 2 connection: Pelle Lykkebo Mørk (8) -- works with data security and has verified E-3 Sentry operations. Understanding AI regulation is directly relevant to his OSINT and data security work.

8

Contradiction Finder

Intermediate

What you will learn

How to use Claude to systematically find inconsistencies between different accounts of the same event.

📄 Suggested document: International Migration — 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.Find two articles about the same event from different news sources. Paste both into Claude.
5.Type: "Compare these two articles. Where do they contradict each other? For each contradiction, quote the relevant passage from both sources."
6.For each contradiction Claude finds, categorize it yourself: factual disagreement, different emphasis, or different sourcing?
7.Ask: "Which version is more likely to be accurate based on the sourcing each article provides? Explain your reasoning."
8.Challenge Claude: "What information would we need to resolve these contradictions?"

What to watch for

Claude may flag differences in emphasis as contradictions, or miss subtle but important factual disagreements. It's better at finding explicit contradictions (different numbers, different timelines) than implicit ones (different framing that implies different causation).

TV 2 connection: Mads Buur Bach (6) -- works with video verification and translation, often comparing multiple source accounts. This technique applies directly to cross-referencing conflicting reports.

9

The FoI Request Generator

Intermediate

What you will learn

How to use Claude to draft thorough Freedom of Information requests that cover the angles you might miss.

📄 Suggested document: Danish Contribution to Rule of Law 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.Describe a story you are working on and what government-held information you need.
5.Type: "Draft an aktindsigt (Freedom of Information) request for this information. Be specific about document types, date ranges, and departments. Use Danish legal language appropriate for a formal request under offentlighedsloven."
6.Review the draft. Then ask: "What am I forgetting to ask for? What related documents might exist that I haven't thought of? What loopholes might the authority use to narrow their response?"
7.Iterate: incorporate the suggestions and ask Claude to produce a final version.
8.Have a colleague or editor review the final request before sending.

What to watch for

Claude knows general FoI principles but may not be current on specific Danish procedural requirements or recent court rulings about offentlighedsloven. The "what am I forgetting" step is where Claude adds the most value -- it's surprisingly good at thinking of adjacent document types you might not have considered.

TV 2 connection: Jakob Hohlmann Villumsen (3) -- works directly with FoI requests and people-finding. This exercise was designed specifically for his workflow.

10

The Chain of Reasoning

Intermediate

What you will learn

How to decompose complex claims into individually verifiable facts -- and how to assess which parts are actually checkable.

📄 Suggested document: A.P. Moller-Maersk 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 complex claim from a politician, press release, or opinion piece -- something with multiple embedded assertions.
5.Paste it into Claude and type: "Break this statement into individual factual claims. List each one separately. For each claim, rate it: Verified (I can confirm this from my training data), Unverified (plausible but I cannot confirm), or Unverifiable (requires real-time data or access I don't have)."
6.Review Claude's breakdown. Did it catch all the embedded claims? Did it miss any implied assertions?
7.For claims rated "Verified," ask: "What is your source for verifying claim #3?" -- can Claude justify its confidence?
8.For "Unverifiable" claims, ask: "What would I need to check to verify this, and where would I look?"

What to watch for

Claude is excellent at decomposing claims but can be overconfident about "Verified" ratings. Its training data is not a reliable source -- treat "Verified" as "plausible based on training data" and verify independently anyway. The real value is the decomposition itself: turning one complex sentence into a checklist.

TV 2 connection: Peter Møller (7) -- focuses on verification and scientific sources. This systematic approach to claim decomposition directly supports his verification workflow.

11

Methodology Critique

Intermediate

What you will learn

How to use Claude as a critical reader of academic research -- identifying weaknesses before you cite a study in your reporting.

📄 Suggested document: Blessing or Curse? AI and Misinformation (arXiv)
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.Upload an academic study or research report relevant to a story you're working on.
5.Type: "Analyze the methodology of this study. What are the methodological weaknesses? What alternative explanations exist for the findings? What important variables are missing or uncontrolled?"
6.Follow up: "If I cite this study in an article, what caveats should I include? What would a critic say?"
7.Ask: "What questions should I ask the authors in an interview?"
8.Compare Claude's critique to what an actual domain expert would say -- where does Claude's analysis hold up, and where does it show gaps?

What to watch for

Claude is good at identifying common methodological issues (sample size, selection bias, confounders) but may miss domain-specific problems that only an expert would catch. Use this as preparation for expert interviews, not as a replacement for them.

TV 2 connection: Pelle Lykkebo Mørk (8) -- with his OSINT and breaking news work, evaluating the reliability of source material is essential. This technique applies to any document, not just academic papers.

12

Meeting Notes to Action Items

Intermediate

What you will learn

How to turn messy, unstructured meeting notes into clear, actionable output -- and where Claude fills in gaps it shouldn't.

📄 Suggested document: Danish Economic Survey — December 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.Paste raw, messy meeting notes into Claude -- the less structured, the better.
5.Type: "Extract from these notes: (1) Decisions that were made, (2) Action items with the person responsible, (3) Open questions that remain unresolved, (4) Any deadlines mentioned."
6.Check every item Claude extracts against the original notes. Did it infer anything that wasn't explicitly stated?
7.Ask: "Which of these items did you infer rather than directly extract? Flag anything that is your interpretation rather than explicit in the notes."
8.Refine the output and save it as a template for future meetings.

What to watch for

Claude has a tendency to infer action items that weren't explicitly assigned and to assign ownership based on who was speaking, not who was tasked. Step 4 is critical -- always ask Claude to distinguish between extraction and interpretation.

TV 2 connection: Emil Gjerding Nielson (8) -- with his editorial AI focus, this technique streamlines the editorial workflow from meeting to execution.

13

The Source Credibility Matrix

Intermediate

What you will learn

How to use Claude to systematically evaluate and compare the credibility of multiple sources on the same topic.

📄 Suggested document: IPCC AR6 Summary for Policymakers
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.Collect 5 sources on the same topic: a news article, a press release, an NGO report, a government statement, and a social media post.
5.Paste all five into Claude and type: "Rank these 5 sources by credibility. For each source, assess: potential bias, funding source or institutional affiliation, track record (if identifiable), and what each source has to gain or lose from their position."
6.Challenge the ranking: "What would change your ranking? What information would make the lowest-ranked source more credible?"
7.Ask: "Which source is most likely to be wrong, and about what?"

What to watch for

Claude tends to favor institutional sources over individual ones, and may rank "official" higher than "credible." It also cannot verify track records in real time -- its assessments of source reliability are based on general patterns, not current reputations. Use this as a structured thinking exercise, not as definitive source evaluation.

TV 2 connection: Peter Møller (7) -- works with verification and scientific sources. This matrix approach systematizes what good journalists do intuitively, making it teachable and repeatable.

14

Translation Quality Check

Intermediate

What you will learn

How to use competing AI translations to catch nuances that a single tool misses -- particularly important for non-English source material.

📄 Suggested document: ECHR Judgment: Lings 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.Take a text in a foreign language (Arabic, Ukrainian, German -- whatever your story requires). Translate it in ChatGPT.
5.Translate the same text in Claude.
6.Paste both translations into Claude and type: "Here are two translations of the same text. Compare them. Where do they differ? What nuances does Translation A capture that Translation B misses, and vice versa? Are there any passages where the meaning changes between translations?"
7.For any significant differences, ask: "Which translation is more accurate for the passage about [specific topic]? What is the original actually saying?"
8.If the text is critical to your story, use the comparison as preparation for speaking to a human translator.

What to watch for

Claude may favor its own translation in the comparison -- which is why you should present both anonymously ("Translation A" and "Translation B") without revealing which tool produced which. Neither AI translation should be considered authoritative for sensitive reporting -- use this as a triage step.

TV 2 connection: Mads Buur Bach (6) -- already works with video verification and translation. This technique adds a systematic quality check to his translation workflow.

Exercises 15 – 20

Advanced

Complex workflows. Build reusable systems and push Claude to its limits.

15

Multi-Document Cross-Reference

Advanced

What you will learn

How to use Claude to find inconsistencies across multiple documents from the same organization -- the kind that reveal what they don't want you to notice.

📄 Suggested document: Carlsberg 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.Collect 3 related documents from the same organization: annual reports from different years, a press release and the underlying data, a public statement and an internal memo.
5.Upload all three and type: "These three documents are from the same organization. Where do they tell different stories about the same facts? Look for: numbers that changed between reports, claims in one document contradicted by another, and information present in one but conspicuously absent from others."
6.For each inconsistency Claude finds, ask: "Is this a genuine inconsistency or a legitimate update/correction? What would explain this change?"
7.Ask: "Based on these documents, what are the three most important questions you would ask this organization?"
8.Create a timeline: "Build a chronological timeline of claims across all three documents. Where does the narrative shift?"

What to watch for

Claude can process multiple documents simultaneously, but its analysis quality decreases as document length increases. If you're working with very long documents, consider uploading key sections rather than full texts. Also watch for Claude finding "inconsistencies" that are actually just different contexts -- not every difference is a contradiction.

TV 2 connection: Mathias Overgaard (5) -- works with OSINT and large reports. Multi-document analysis at this scale is exactly where Claude saves hours of manual cross-referencing.

16

The Bias Detector

Advanced

What you will learn

How to systematically identify rhetorical techniques, hidden assumptions, and missing context in political communication.

📄 Suggested document: European Parliament AI Briefing 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 political speech, press conference transcript, or opinion piece.
5.Paste it into Claude and type: "Analyze this text for: (1) Every instance of framing -- how word choices shape perception, (2) Loaded language -- emotionally charged words that could be replaced with neutral alternatives, (3) Unstated assumptions -- what the speaker takes for granted, (4) Missing context -- what information is left out that would change the interpretation."
6.For each instance of framing, ask Claude to provide the neutral alternative.
7.Ask: "If you were editing a news article based on this speech, what would you remove, what would you keep, and what context would you add?"
8.Test Claude's own bias: "Are you being harder on this text because of its political position? Analyze your own analysis for bias."

What to watch for

Claude has its own biases -- it tends to be more critical of populist rhetoric and more forgiving of technocratic language. Step 5 is essential: make Claude examine its own analysis. Also note that "loaded language" is context-dependent; a word that's loaded in one context may be precisely accurate in another.

TV 2 connection: David Buch (7) -- strong critical thinker. This exercise gives his critical instincts a systematic, repeatable framework that can be applied to any political communication.

17

Prompt Engineering Lab

Advanced

What you will learn

How to build, test, and iterate on reusable prompt templates -- turning ad hoc requests into reliable tools.

📄 Suggested document: EU Rule of Law Report — 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.Type: "Write a reusable prompt template for analyzing Danish parliamentary questions (Folketingets spørgsmål). The template should work for any question, extract the key claim, identify the minister being questioned, and assess whether the question is genuine inquiry or political positioning."
5.Test the template on a real parliamentary question. Document what works and what doesn't.
6.Tell Claude what failed: "The template missed X and misidentified Y. Revise it."
7.Test the revised template on a different parliamentary question. Iterate at least 3 times.
8.Save the final template with a note about what it's good at and where it still needs human judgment.

What to watch for

The first version of any prompt template is never good enough. The value is in the iteration. By version 3, you'll have a prompt that handles edge cases the first version missed. Keep all versions -- the evolution shows you how Claude interprets instructions and where ambiguity causes problems.

TV 2 connection: Sanne Lau Pedersen (5) -- interested in monitoring and automation, but Python didn't work for her. Prompt engineering is a more accessible path to automation that doesn't require coding.

18

The Investigative Thread

Advanced

What you will learn

How to use Claude to generate investigative leads from public financial documents -- the questions that numbers raise but don't answer.

📄 Suggested document: Danske Bank 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.Upload a company's public filings, annual report, or financial statements.
5.Type: "You are an investigative journalist reviewing these financial documents. What questions would you ask about these numbers? Look for: unusual changes year-over-year, items that seem disproportionate, related-party transactions, and anything that seems designed to obscure rather than clarify."
6.For the most interesting questions, ask: "What public records or additional documents would help answer this question?"
7.Ask: "If these numbers are legitimate, what innocent explanations exist? If they're not legitimate, what might be happening?"
8.Create a research plan: "Prioritize these questions. Which ones are most likely to lead somewhere, and what's the most efficient order to investigate them?"

What to watch for

Claude can generate investigative questions but cannot verify whether those questions lead anywhere. It may flag things as suspicious that have perfectly innocent explanations. Treat the output as brainstorming, not as findings. The value is in generating leads you wouldn't have thought of, not in drawing conclusions.

TV 2 connection: Franziska Weiss Lauritzen (5) -- works in criminal journalism. Following the money is fundamental to her beat, and this technique helps generate leads from public documents before investing weeks of reporting.

19

The Ethical Dilemma Workshop

Advanced

What you will learn

How to use Claude as a thinking partner for ethical decisions -- testing arguments from both sides before committing to a position.

📄 Suggested document: EU Surveillance — Denmark (FRA)
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.Present a real journalism ethics dilemma to Claude. Examples: publishing a leaked document that could endanger a source, naming a suspect before charges, using AI-generated images in news coverage.
5.Type: "Argue the case FOR publishing/proceeding. Give me the strongest possible argument, citing journalistic principles and precedent."
6.Then: "Now argue the case AGAINST. Same rigor, same depth."
7.Ask: "Which argument is stronger, and why? What is the single most important factor in this decision?"
8.Challenge: "What would a Danish media ethics expert say about your conclusion? What would Pressenævnet consider?"

What to watch for

Claude has a built-in caution bias -- it tends to argue more strongly against publication and in favor of caution. This doesn't mean caution is always right. Be aware of this tendency and push back when the "publish" argument is genuinely stronger. Also note: Claude's knowledge of Pressenævnet rulings may be limited or outdated.

TV 2 connection: Pelle Lykkebo Mørk (8) -- his data security work regularly involves ethical decisions about what to publish. This exercise builds a structured framework for those decisions.

20

Build Your Beat Archive Prompt

Advanced

What you will learn

How to design a reusable, beat-specific Claude prompt that works across different stories -- your personal AI analysis tool.

📄 Suggested document: Pandora A/S 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.Define your beat: What kind of documents do you regularly analyze? What questions do you always ask? What patterns do you always look for?
5.Tell Claude: "I cover [your beat] for TV 2 Fyn. Help me design a reusable prompt that I can use every time I receive a new [document type]. It should extract [key information], flag [common red flags], and generate [useful outputs]. The prompt should work regardless of the specific document."
6.Test the prompt on a real document from your beat. Note where it works and where it doesn't.
7.Test it on a second, very different document from the same beat. Does it still work?
8.Test it on a third document. Iterate until the prompt handles the variety of your beat. Save the final version.

What to watch for

A prompt that works perfectly for one document often fails on the next. The challenge is finding the right level of generality -- too specific and it only works for one type; too general and it produces shallow analysis. The best beat prompts have a fixed structure (what to always check) with flexible sections (what varies by document).

TV 2 connection: Lars Apel (7) -- interested in agents and news scanning. This exercise is the foundation for building automated monitoring prompts. Sanne Lau Pedersen (5) -- this is the non-code version of the automation she was trying to build with Python.

When to use Claude -- and when NOT to

Use Claude when

  • You need to analyze a long document and extract specific information with references
  • You want honest assessment of what's verifiable and what's not in a claim
  • You need to compare multiple documents and find inconsistencies
  • You want to brainstorm investigative questions or FoI angles
  • You need structured extraction -- tables to CSV, notes to action items, claims to checklists
  • You want to test arguments from multiple sides before taking a position
  • You need a critical reader for academic studies or legal documents

Do NOT use Claude when

  • You need current information -- Claude has no internet access and a knowledge cutoff
  • You need verified facts -- Claude's training data is not a source, it's a starting point
  • You need to search the web, check a URL, or verify that a website exists
  • You need a direct quote from a person -- Claude cannot verify quotes
  • You're writing a final article without human editing -- Claude writes like Claude, not like you
  • You need to verify images or video -- Claude can describe what it sees but not authenticate
  • You need legal certainty -- Claude can identify relevant law but not interpret it authoritatively

The golden rule: Claude is your research assistant, not your editor-in-chief. It reads faster than you and misses different things than you do. The combination -- your judgment plus Claude's speed -- is more powerful than either alone. But your judgment is the one that matters.