Digital Digging × AI Training for TV 2 Journalists

The Prompt Behind
the Neutral Report

In 8 parts — how words steer the machine

April 2026 20 Participants Meta-document

Most people think AI “just knows” what to do. You type a question, you get an answer. But everything — the tone, the structure, what is included and what is left out — depends on the instruction. The prompt. The neutral report you have seen looked objective. Dry. Factual. But behind that apparent neutrality sat a carefully constructed instruction of hundreds of words, divided across 8 parts.

This document dissects that prompt. Each part is shown as it was actually given to the AI, with an explanation beside it: what does this part do, and why is it there? This makes visible how language — very precisely chosen language — determines what a machine produces.

1

Instruction

The opening task given to the AI

Analyse the attached intake forms of 20 journalists at TV 2 Odense. These are staff members participating in an AI training delivered by Digital Digging (Henk van Ess). Generate a NEUTRAL analysis report based on this data. This report must be factual. It is one of three variants (CRITICAL, POSITIVE, NEUTRAL) that together serve as teaching material about how framing works.

What this part does

The first sentences determine everything. They give the AI three frames: who the audience is (journalists, not engineers), what the goal is (a neutral report), and what context it sits in (part of a triptych about framing). Without this opening the AI would choose its own tone — and that choice would be invisible.

Notice the word “NEUTRAL” in capitals. That is not shouting — it is an anchor. The AI weighs words in capitals more.

2

Source Data

Reference to the CSV input

The CSV file contains 20 entries with the following columns: name, self-assessed AI proficiency (scale 1–10), current AI usage (open answer), goals (open answer), success criteria (open answer), and additional comments. The scores range from 3 to 9, with a group mean of 6.0. Treat each entry individually — do not aggregate to group level before reporting on each person first.

What this part does

This part tells the AI exactly what data is available and how it is structured. Without this description the AI would have to guess which columns are relevant.

The last sentence is the hinge: “Treat each entry individually.” Without it the AI tends to average and generalize immediately — individual answers disappear into group statistics. The order (individual first, then group) is a deliberate choice.

3

Assignment

The specific analysis task

Write a report that states what participants filled in — not more, not less. Use their own words where possible. Paraphrase only when the original is unclear. Where the CRITICAL variant emphasizes problems and the POSITIVE variant sees opportunities, the NEUTRAL variant does neither. It states. It describes. It does not interpret. Example: if someone writes "I still find googling stuff myself faster", report: "Participant states that manual search is sometimes faster than AI." Not: "Participant resists new technology" (critical). Not: "Participant has high editorial standards" (positive).

What this part does

Here the core of the neutral approach is defined — not by saying what the AI must do, but by showing what the other two variants do and then saying: do neither.

The example at the end is essential. AI models respond strongly to concrete examples. One example is more effective than ten lines of abstract instruction. It lets the AI feel the difference between stating and interpreting.

Notice that the example uses Lars Apel’s actual words from the intake. Using real data in the instruction anchors the AI to this specific dataset.

4

Report Structure

How the report must be built

Use this structure: 1. HEADER — Title, context, badges (date, number of participants, report type) 2. KEY FIGURES — 5 cards: mean score, median, score range, number of participants, most mentioned need 3. SCORE DISTRIBUTION — Visual display of all 20 scores 4. PER PARTICIPANT — Two columns: "Current State" (what they do now) and "Identified Gaps" (what they themselves name as missing) 5. THEME ANALYSIS — Which topics recur, with frequency 6. RAW DATA — Direct quotes from the forms 7. METHOD — How the report was produced

What this part does

Structure is framing. By choosing two columns (“Current State” and “Identified Gaps”) you already make a choice: you separate what someone can from what someone lacks. That appears neutral, but it steers the reader toward a deficit perspective.

An alternative would have been: one column with “Summary of answers.” No separation. Less clear, but also less steering. This is the kind of choice that remains invisible when you don’t know the prompt.

5

Writing Rules

Rules for neutral writing

Writing rules for the NEUTRAL variant: • No value judgments. Not "good", not "bad", not "notable". • No superlatives. Not "the most", "the best", "the largest". • No advice. Not "could benefit from", "it is recommended", "an opportunity". • No emotional language. Not "unfortunately", "fortunately", "surprisingly". • No comparisons between participants. Not "in contrast to" or "just like". • Use: frequencies ("7 of the 20 participants mention..."). • Use: direct quotes with quotation marks. • Use: descriptive statistics (mean, median, range). • Tone: as if you are writing an inventory list. Dry. Precise.

What this part does

This is the longest and most detailed section — and that is telling. Neutral writing is harder than colored writing, also for an AI. Without these explicit prohibitions, value judgments creep in by themselves.

Notice the pattern: the rules first state what is not allowed (five prohibitions), then what is allowed (three permissions). This pattern — prohibition-before-permission — is more effective than permission alone. The AI knows exactly where the boundaries are.

The last line (“as if you are writing an inventory list”) is a metaphor. AI models understand metaphors surprisingly well — better than abstract instructions.

6

Comparison

How three variants frame the same data differently

For reference: below is how the same participant data is formulated differently in each of the three variants. Use this as a benchmark to verify that your NEUTRAL formulation actually holds the middle ground — stating, without charge.

The power of framing becomes visible when you place the same data side by side. Below five participants — the same answers, three completely different interpretations:

Participant
Critical
Positive
Neutral

Jakob

Score: 3

“I spend way too much time on tracking down people by old school browsing on Facebook”

“Still browsing Facebook manually. Score 3 reflects a workflow that has not evolved beyond basic social media browsing.”
“Has the most specific investigative use case in the group. Already connecting dots across profiles — AI will accelerate this.”
“Self-assessment: 3/10. Uses ChatGPT for FoI requests and verdict summaries. States time is spent on manual social media research.”

Lars

Score: 7

“I still find googling stuff myself faster than using AI”

“Agent experiments failed. Despite score 7, defaults to Google. AI adoption is surface-level.”
Pushing real boundaries by attempting agents. Honest assessment shows editorial maturity.”
“Self-assessment: 7/10. Reports attempted agent setup for news scanning. States manual search is sometimes faster than AI.”

Christian

Score: 9

“I am very skeptical with AI as a tool for generating knowledge”

“A 9/10 who does not trust the technology he rates himself on. Contradiction between score and stated skepticism.”
Deepest technical understanding in the group. Skepticism reflects informed use, not resistance.”
“Self-assessment: 9/10. Uses AI for summarization. States skepticism about AI for knowledge generation. Notes language models are suited for text tasks.”

Sanne

Score: 5

“I used Phyton and Task Scheduler — but it never worked”

“Tool ambitions exceed current skills. Python attempt failed. Gap between aspiration and execution.”
Most technically ambitious participant. Attempted real automation. Ready for the right tools.”
“Self-assessment: 5/10. Reports attempted Python-based monitoring using Task Scheduler. States the setup did not succeed.”

Pelle

Score: 8

“the risk of offloading independent editorial judgement to AI”

“Admits to dependency risk. Speed is replacing judgment. Awareness alone does not prevent the pattern.”
“Shows rare self-awareness about AI limitations. Already verified E-3 Sentry damage via flight tracking data.”
“Self-assessment: 8/10. Uses AI across research, drafting, and live production. Notes concern about editorial judgment being displaced by speed.”

The lesson from this table

The same skepticism from Christian is read by the critical variant as a contradiction, by the positive variant as informed expertise, and by the neutral variant simply as a stated position. The same failed Python script from Sanne becomes a “skills gap,” a “technical ambition,” or a reported outcome. The data does not change. The words around it change everything.

This is why the prompt matters. Not the data, but the instruction determines the story.

7

Format Instructions

Technical layout: HTML, Tailwind, colors

Generate the report as one complete, self-contained HTML file. Use Tailwind CSS via CDN. Use the following visual style: Fonts: DM Sans (text), JetBrains Mono (labels, code) Colors: paper #f7f5f0, cream #ede9e0, ink #0a0a0a, smoke #6b6560, ash #9c9590, rust #c44b28 Maximum width: 820px, centered Animations: fadeUp on scroll Texture: SVG noise overlay No external images. No JavaScript frameworks. No build steps. The file must open directly in a browser.

What this part does

Design is also framing. The choice of warm, muted colors (#f7f5f0 paper, #6b6560 smoke) gives the report a calm, almost academic feel. A report with the same content in bright red and yellow would feel completely different.

The technical requirement “one self-contained HTML file” is practical: the report can be opened on any computer, without installation. But it is also a choice for accessibility — nobody needs a developer to read it.

8

Control

Verification checklist

Verify before delivery: ✓ Are all 20 participants individually included? ✓ Are the calculations correct (mean, median, range)? ✓ Does the report contain zero value judgments? (search for: "good", "bad", "strong", "weak", "unfortunately", "fortunately") ✓ Are quotes rendered verbatim, with quotation marks? ✓ Is there no advice or recommendation anywhere? ✓ Does the word "should" not appear? ✓ Is the HTML valid and does it open without errors in a browser? ✓ Is there a method section at the bottom explaining how the report was produced?

What this part does

This is the safety-net section. Without explicit verification requirements an AI model sometimes produces output that meets most criteria, but deviates subtly — a single “unfortunately” that slips in, a participant who gets skipped, a mean that doesn’t add up.

The search words (“good”, “bad”, “strong”, “weak”) function as a negative word list. By explicitly asking the AI to search for those words, the chance that they appear in the output decreases. It is self-correction, built into the assignment.

Conclusion

The Lesson

The neutral report appears the most “honest” of the three variants. It does not judge. It does not advise. It states. But even that apparent objectivity is the result of choices — choices made in the prompt, not by the data.

Three choices that make “neutral” still steering:

01

The two-column structure — By placing “Current State” and “Identified Gaps” side by side, the report suggests that something is missing. A one-column summary would not create that suggestion.

02

The choice for frequencies — “7 of the 20 participants mention...” sounds neutral, but it implicitly tells you that 13 participants do not mention it. That is also framing.

03

The concept “gaps” — The word “gap” sounds neutral but presupposes that something is missing. An alternative like “stated wishes” would present the same data without a deficit connotation.

True neutrality does not exist. What does exist is transparency about your choices. This document is that transparency: it shows which instructions were given, which were not, and why. That is ultimately worth more than the illusion of objectivity.

Accountability

What Was Deliberately Left Out

The prompt for the neutral report is not only defined by what is in it, but also by what is deliberately not in it. Every omission is a choice.

No instruction to compare participants with each other

Comparison inevitably creates a hierarchy. “Bo scores higher than Jakob” sounds factual, but positions Jakob as “less.” By leaving out comparison, each participant stands on their own.

No instruction to suggest tools or solutions

The moment you name a tool (“Claude could help here”), an inventory becomes a sales pitch. The neutral report describes what people themselves name, not what the AI thinks they need.

No instruction to predict success or failure

Predictions are by definition speculative. “With these scores the training will probably...” is not reporting but fortune-telling. A neutral report contains exclusively what is, not what might come.

No instruction to group into “good” and “bad”

The critical variant implicitly divides the group into leaders and laggards. The positive variant into starters and advanced users. The neutral report refuses every classification — because every classification is a judgment.

No instruction to create urgency

Words like “crucial,” “urgent” or “act now” are powerful framing instruments. They steer the reader toward the conclusion that something must change. The neutral report leaves that conclusion to the reader.

What Was Actually Typed

Two Commands. That Was It.

You have just read 8 sections full of detailed instructions, writing rules, verification requirements and structural prescriptions. Hundreds of words in total, carefully constructed. But what was actually typed into the AI’s input field was this:

Step 1 — verbatim

“Give me a prompt to analyse (subject)

Step 2 — verbatim

“Execute prompt”

Two commands. Nine words total. That was it.

How is that possible?

This is the Prewash Method. In Step 1, the CSV file and instruction file were uploaded alongside the command. The AI read the structure of the data and the instructions, and designed its own analytical prompt — complete with writing rules, structure, verification criteria, everything you read in the 8 parts above.

In Step 2, the AI was told to execute the prompt it had just generated. The result was the neutral report — not from a hand-typed 500-word instruction, but from a two-step process where the AI itself determined the analytical framework.

The difference with a direct command (“Summarise this data”) is fundamental. A summary gives you an echo of the author. The Prewash Method gives you an analysis from angles you had not thought of, because the AI reads the structure of the document and asks it questions you would not have formulated yourself.

The lesson for your work

When you use AI for newsroom work at TV 2, you do not need to write elaborate prompts. Upload your document. Type “Give me a prompt to analyse (subject).” Read what the AI proposes. Adjust if needed. Then type “Execute prompt.”

That is the core skill for all work with AI: you don’t know what AI can ask a document. The AI reads structures, patterns and gaps that you as a human would not spot on first reading. By letting the AI design the prompt, you get analysis instead of summary. You get questions instead of answers. And that is exactly what journalism needs.