---

The example: university or apprenticeship?

I ran the prompt on a question every graduating class in Germany faces: should you go to university in 2026 or do an apprenticeship? And if so, which one?

The result is a research report with 54 sources, 24 of them official statistics from the Federal Employment Agency, BIBB, the OECD and others. The verdict for a concrete profile (Abitur grade 2.3, into tech and business, wants to earn money soon): dual study program, medium confidence.

Three findings that surprised me:

• A degree pays more but barely protects you more. Among 25 to 34 year olds, 3.5 % of graduates are unemployed versus 3.3 % of people with vocational qualifications (OECD, Education at a Glance 2025).
• The bigger risk is dropping out. At universities, 42 to 47 % drop out of computer science and 44 to 51 % out of electrical engineering (DZHW). Apprentices with an Abitur end only 17.9 % of their contracts early (Vocational Training Report 2026).
• Young graduates in Germany show no measurable AI effect through the end of 2025. The negative effect shows up in non-academic jobs and manufacturing, and there the economic cycle hits more than twice as hard (IAB-Forum, not yet peer reviewed).

Here is the full research as a page (in German): verdict, rankings of degree programs and apprenticeships, a matrix for other profiles, AI exposure by job, and all 54 sources, clickable.

This is not career advice. The research covers one example profile, and salary portals disagree by several thousand euros in places. The page flags those spots.

Step 1: Adapt the prompt to your question

The prompt below is built for the job-market question. For a different question, swap three parts and keep the rest:

• Context: who you are, what you want to decide, what resources you have. The more concrete, the clearer the verdict. With a vague profile, even this prompt answers "it depends".
• Scope: countries, time frame, which options get compared.
• Phases 2 to 5: the subject-matter questions. Phase 1 (hypotheses) and phases 6 to 8 (synthesis, bias check, verdict) stay as they are.

If rewriting it feels fiddly, let Claude do it:

Here is a research prompt for Perplexity. Adapt it to my question.

My question: [What do you want to decide?]
My context: [Who you are, what's already fixed, budget, time]
Countries and time frame: [e.g. US, data from 2022]

Keep the Role, Phase 1, the synthesis, bias check and verdict phases,
the quality rules and the output format. Replace Context, Scope and
the subject-matter phases. At least two hypotheses must argue against
my obvious answer.

[paste the prompt from below]

Read the result once before it goes to Perplexity. The counter-hypotheses matter most: if they're too soft, the research ends up just agreeing with you.

Step 2: Run it in Perplexity

1. Open Perplexity and switch to a deep-research mode. I use Research mode, Labs works too.
2. Paste the prompt and send it.
3. Let it run. For me a run takes a few minutes.
4. At the end you get two files: a JSON with all the data and a Markdown summary. Download both.

Quick sanity check: look at the source list. For the job-market question it should include the Federal Employment Agency, IAB, BIBB and DZHW. If most primary sources are missing and it's all career portals, the research was too shallow. Run it again.

Step 3: Claude turns it into a page

Drop both files into a new Claude chat and add:

Here are the results of my research: the summary and the JSON with
all the data. Build an artifact from it:

- Verdict at the top, with confidence and who it applies to
- The core finding as a chart
- Then rankings, hypotheses, bias check, open questions
- Every number links to its source
- Forecasts, assumptions and contradictions clearly flagged

Rule: every number on the page must appear verbatim in the JSON.

The last line matters most. Without it, Claude will sometimes round a number or merge two values while designing. With it, the page stays as well sourced as the research.

If you don't like something, just say so in the chat ("rankings as a table", "bigger chart", "dark theme"). Claude rebuilds the page.

Step 4: Share it

The artifact starts out private. Use the Share menu on the page to publish it and get a link you can pass on. People see the finished page, not your chat.

The prompt

This is the exact prompt behind the job-market research. Its last quality rule makes the output German; change that line to get English.

Role
You are a senior labor-market economist and career researcher
specializing in the German-speaking labor market, vocational education
and the impact of AI on jobs. Work like a top-1% researcher: frame the
question sharply (Curie), actively look for the opposing view and
check for bias (Kahneman), connect disciplines such as labor economics,
education research, demography and AI exposure research (Da Vinci), and
state conclusions with ruthless clarity (Newton). The goal is truth,
not validation.

Context
I am 18, finishing my Abitur in Germany in 2026 (average grade 2.3).
I am interested in technology and business but have no fixed passion
yet. My parents can support me modestly; I would like to earn my own
money soon. I want to decide: university degree or vocational training
(Ausbildung / Lehre)? If university, which field? If vocational
training, which occupation? Your research is the basis for my decision.

Scope
* Regions: Germany first, Austria and Switzerland as comparison
* Time: data from 2022 onward, labor-market outlook to 2035
* Paths: university (Bachelor/Master), dual vocational training,
  dual study programs (duales Studium) as hybrid, further training
  after vocational training (Meister, Techniker, Fachwirt)

Process (run the phases strictly in order)

Phase 1: Frame the question
1. Write 5 SMART hypotheses. At least two must argue against a
   university degree, e.g. "Vocational training plus Meister/Techniker
   reaches lifetime earnings within 10 % of a Bachelor degree in
   technical fields."
2. Derive 3 sharp test questions per hypothesis.

Phase 2: Labor market 2026
3. Collect hard data: unemployment rate by qualification level,
   shortage occupations, open positions vs. applicants, entry salaries
   and lifetime earnings by qualification, dropout rates (university
   and vocational training).
4. Separate short-term signals (2025/2026 economy) from structural
   trends (demography, retirement of baby boomers, energy transition).

Phase 3: AI and entry-level jobs
5. Which entry-level tasks and occupations are most exposed to AI,
   which are most resilient (manual, physical, regulated, people-facing
   work)? Back each with a study. Separate evidence from forecasts.

Phase 4: If university, then what
6. Rank max. 8 fields of study for my profile by: employment rate,
   entry salary, AI resilience, demand outlook to 2035, dropout risk.
   Give numbers with sources for each criterion.

Phase 5: If vocational training, then what
7. Rank max. 8 vocational occupations (Ausbildungsberufe) for my
   profile by the same criteria, plus training allowance and paths to
   further training. Include at least 2 dual study programs as hybrids.

Phase 6: Synthesis
8. Build a decision matrix: which person profile fits which path
   (grades, interests, risk appetite, financial situation, preference
   for theory vs. practice). Mark contradictions.

Phase 7: Bias audit
9. Name 3 ways this research could mislead me (e.g. averages vs. my
   individual case, past earnings data vs. future AI effects, survivor
   bias in success stories).
10. Devil's advocate: give the strongest argument for the path you do
    NOT recommend.

Phase 8: Verdict
11. One verdict for my profile: STUDY / APPRENTICESHIP / DUAL_STUDY,
    with a confidence level (high / medium / low), plus the top 3 picks
    of the chosen path and the conditions under which the verdict flips.

Quality rules
* Every number and claim needs a source (URL, publisher, date).
* Forecasts and projections are labeled "forecast", working papers and
  non-peer-reviewed studies "preprint".
* If you cannot find evidence, label the point "assumption" or "open
  question". Never state it as fact.
* Where sources conflict, show both and mark it CONTRADICTION.
* Primary sources (Bundesagentur für Arbeit, IAB, BIBB, Destatis, DZHW,
  OECD) before press articles and career portals.
* Write all output in German.

Output
Deliver two files.
1. `jobmarkt-2026-research.json`
{
  "meta": { "created": "", "scope": "", "method": "" },
  "verdict": { "decision": "STUDY|APPRENTICESHIP|DUAL_STUDY",
    "confidence": "", "rationale": "", "top_picks": [],
    "flip_conditions": [] },
  "hypotheses": [ { "id": "", "hypothesis": "",
    "status": "confirmed|rejected|open", "test_questions": [],
    "findings": [ { "claim": "", "value": "", "source_id": "" } ] } ],
  "labor_market": [ { "indicator": "", "value": "", "year": "",
    "trend": "short_term|structural", "source_id": "" } ],
  "ai_exposure": [ { "occupation_or_task": "",
    "exposure": "high|medium|low", "evidence": "",
    "evidence_type": "study|forecast", "source_id": "" } ],
  "study_options": [ { "rank": 0, "field": "", "employment": "",
    "entry_salary": "", "ai_resilience": "", "outlook_2035": "",
    "dropout_risk": "", "source_ids": [] } ],
  "apprenticeship_options": [ { "rank": 0, "occupation": "",
    "type": "apprenticeship|dual_study", "training_allowance": "",
    "employment": "", "entry_salary": "", "ai_resilience": "",
    "outlook_2035": "", "further_training": "", "source_ids": [] } ],
  "decision_matrix": [ { "profile": "", "recommended_path": "",
    "why": "", "evidence_ids": [] } ],
  "bias_audit": [ { "risk": "", "explanation": "" } ],
  "devils_advocate": "",
  "open_questions": [],
  "sources": [ { "id": "", "title": "", "publisher": "", "date": "",
    "url": "", "type": "statistic|study|forecast|preprint|press" } ]
}
2. `jobmarkt-2026-summary.md`
Readable in 60 seconds:
* Headline (max. 12 words) stating the verdict
* 3 key insights, each with a number and a source
* My top 3 options with one line each
* When I should choose the other path
* One next step for this month
Tone: sharp, plain, no jargon, no emojis.

Common mistakes

• Only hypotheses that back your opinion. Then the research just confirms you. At least two have to argue against, and you need to read them yourself.
• No concrete profile in the Context. "I'm wondering what to do after school" gets you an overview. "18, grade 2.3, wants to earn money soon" gets you a verdict.
• Reading forecasts as facts. What happens by 2035 is a scenario. The prompt has forecasts flagged, and the page shows them differently. For important numbers, check whether "forecast" is attached.
• Cutting the output section. Without a fixed JSON you get a long report, and Claude has to pull the numbers back out. That's where errors creep in.
• Dropping the check rule in step 3. Then a number can get rounded or invented during design without you noticing.

Next steps

1. Open the job-market research and look at how the page is built.
2. Pick a decision you're facing right now and adapt the prompt with step 1.
3. Run it in Perplexity and hand the two files to Claude.