Apply Program Theme Extractor

Program Theme Extractor

Before anyone can ask "does my work fit here", someone has to answer "what is here". This skill answers only the second question, for one target or for a whole candidate list, with no applicant evidence in play.

A programme's theme is published selection content, not marketing. Where it is declared, quote it. Where it is not declared, it may still be operative — visible in what the programme teaches, funds, and graduates — or the programme may be genuinely open, which is an equally real answer.

The load-bearing distinction in this skill:

• A declared theme is what the institution says it selects for. It can disqualify a target.
• An inferred theme is your hypothesis about what it selects for. It can shift emphasis. It can never disqualify a target.

Presenting an inferred theme with the authority of a declared one is worse than not running the screen at all — it manufactures a selection criterion that nobody published.

Use this when

• The user has one target and its focus is unclear from the landing page.
• The user has a candidate list (a consortium's pathway list, a doctoral school's research areas, a country's advertised positions, a saved shortlist) and wants each characterised before comparing anything to their profile.
• The user asks whether a call is topic-restricted or open.
• The user needs a pathway-level answer for a multi-specialisation joint programme.
• A downstream skill needs a theme dossier that does not exist yet.

Do not use this when

• The applicant's evidence is in play and the question is fit. Use apply-program-fit-mapper — it owns the core / adjacent-bridged / stretch / outside-scope verdict and the swap test.
• The question is about one professor's research direction. Use apply-research-direction-mapper.
• The question is how a target admits or funds people rather than what it selects for thematically. That is the application-regime and funding-regime axis inside apply-program-fit-mapper.
• The user wants statement prose built on a theme. Use apply-sop-writer.

Workflow

1. Fix the unit of extraction

A theme attaches to a specific selecting body. Before searching, decide which one the user is asking about, because a programme and its pathways often differ:

• a whole programme or consortium
• one specialisation, track, or pathway inside it
• a doctoral school, graduate school, or research centre
• one funding call or cycle
• one advertised position or project

For multi-pathway programmes, extract at the pathway level as well as the umbrella level. Fitting the umbrella while fitting no individual pathway is a common disqualifier, and it is invisible if you only extract the umbrella.

2. Fetch the official pages — always live, never from memory, in two passes

Search the web for every target that is not already in the conversation as page text. A programme you have never heard of is the normal case, not a reason to screen from recall. Titles are marketing and themes change per cycle, so neither the programme's name nor your prior knowledge of the field is evidence.

Inference is search-expensive and the cost scales with the candidate list, so run it in two passes and never spend the expensive one on targets that are already out.

Pass A — cheap, every target. Fetch only the programme or consortium page (plus each pathway page) and the current cycle's call, admission page, or vacancy text. This answers one question: is a theme declared here or not? Targets that come back declared are finished — go to step 3 and stop. Targets that come back link needed are finished too.

Pass B — expensive, only where pass A found nothing declared and the target is still a live candidate. Ask the user which targets survive their other filters (country, funding, degree level, deadline) before spending here; inferring themes for fifteen programmes when four survive is the main waste this skill can cause. Then fetch, recording URL plus access date for each:

• the stated objectives, mission, or strategic-plan page behind the programme
• the funder's own mission page where the programme is funded for an agenda
• module or course lists — compulsory before elective, since required modules define scope and electives advertise breadth
• the thesis or dissertation repository, or any published list of past or example project titles
• the published research-area listing, and the faculty or partner roster as a distribution, not as individual profiles

Where the pages cannot be located, or the cycle's call is unpublished, output link needed and leave the target unclassified. Do not substitute a guess.

3. Extract the declared theme, if there is one

Take the institution's own wording verbatim, with its source. Then decompose it into the concrete research activities it actually covers. Umbrella phrases — "digital and green transition", "smart society", "sustainable industry" — are containers, not topics; pathway descriptions, module lists, partner composition, and example projects show what the container holds. Where the container is genuinely broad, record it as broad rather than narrowing it.

4. Where no theme is declared, test for an operative one

Only after step 3 finds nothing declared, and only for targets that reached pass B. Read the programme's behaviour off its official pages, in tiers, strongest first:

• Tier 1 — funded and advertised work. Project portfolio, advertised positions, funded theses, centre grants. What it pays for is what it selects for.
• Tier 2 — curriculum and output. Compulsory modules first, then specialisation options, past or example thesis titles, graduate destinations.
• Tier 3 — people composition, in aggregate only. The published research-area listing, or the faculty/partner roster read as a distribution, and only where that roster is narrow and published as the programme's scope.
• Tier 4 — framing language. Repeated vocabulary across objectives, news, and admission pages. The weakest tier and the easiest to over-read.

Three gates. An operative theme must pass all three, or the answer is genuinely open.

1. Two sources from two different tiers. A module list and an objectives page that paraphrases that curriculum are one source counted twice. At least one of the two must come from tier 1 or tier 2 — tier 3 and tier 4 cannot establish a theme between them. Record URL and access date for each.
2. Recency. Weight the current cycle's pages, and the last three to five years of projects, positions, and theses. A ten-year portfolio can encode a scope the programme has already left. Where the converging evidence is older than that, say so and downgrade to genuinely open rather than reporting a stale theme as current.
3. The exclusion test — the one that matters most. State what the hypothesised theme rules out. If the scope excludes no plausible applicant — "applied mathematics with societal impact", "data science for a changing world" — it is not a theme, it is a description of the field, and the honest status is genuinely open. An inferred theme that excludes nothing will otherwise fire on almost every target, which is exactly how this skill would decay into the framing generator its status labels exist to prevent.

Stop as soon as all three gates pass. Do not keep fetching to raise confidence in a label that already cannot drop a target. If the gates have not passed after the pass-B page list is exhausted, the answer is genuinely open — not a wider search.

4b. What is not evidence about a programme's theme

• An individual professor's research. It is their direction, not the institution's selection scope, and generalising from three faculty pages invents a criterion nobody published. Person-level direction belongs to apply-research-direction-mapper. Only the aggregate roster distribution in tier 3 counts here.
• The country. A nation has no theme. A country enters only where it produces a documented institutional mechanism, and only through the falsifiable chain geopolitical objective → programme objective → published selection criterion — which is already captured when step 2 fetches the funder's own mission page. "Programmes in this country tend to emphasise X" is a cultural prior, not published selection content, and must never appear in a dossier.
• The programme's name, its department, its ranking, or your prior knowledge of the field.

5. Assign exactly one status, and state it in the output

genuinely open and link needed must both remain visible outputs. Producing a theme for every target turns this skill into a framing generator — precisely the failure the whole screen exists to prevent.

6. Hand off

State explicitly which downstream question this dossier does not answer: whether the applicant fits. Route to apply-program-fit-mapper with the dossier, and carry the status label with it — the fit verdict's power depends on it.

Output format

One block per target.

# Theme Dossier: [Programme / pathway / call]

- **Unit extracted:** [programme | pathway | doctoral school | call | position]
- **Status:** declared | inferred | genuinely open | link needed
- **Cycle:** [year / call reference, or "not published"]

## Sources
| Source | URL | Accessed | What it contributes |
|---|---|---|---|

## Declared theme
> [verbatim institutional wording, or "none published"]

## Decomposition — research activities actually covered
- [Activity / method / domain] — evidence: [which source shows it]
- ...
- Genuinely broad in: [where the container was not narrowed]

## Operative theme (only when status = inferred)
- **Hypothesis:** [one sentence]
- **Converging sources:** [source 1, tier N] + [source 2, tier M] — different tiers, at least one from tier 1 or 2
- **Recency:** [dates of the evidence relied on]
- **Excludes:** [what this scope rules out — if nothing, the status is `genuinely open`, not `inferred`]
- **Not published by the institution** — may inform emphasis, must never be quoted as the programme's declared scope, and cannot drop this target.

## Pathways (multi-pathway targets)
| Pathway | Declared scope | Status |
|---|---|---|

## Hard constraints found
- [topic restrictions, eligible research areas, mandatory specialisation choice, or "none found"]

## Not answered here
- Applicant fit. Route to `apply-program-fit-mapper` with this dossier and its status.

For a candidate list, precede the blocks with a one-row-per-target summary table (target, unit, status, one-line scope) so the user can triage before reading.

Quality bar

A strong output lets the user say what each target selects for, and cite where that came from. Every theme traces to a fetched page with a date. No target carries a scope the institution did not publish without an inferred label attached, and no target is dropped on an inferred theme. Targets whose pages could not be found are visible as link needed rather than quietly missing.

Two counts are worth watching as a self-check. If inferred is firing on most targets, the exclusion test is not being applied and the dossiers are describing fields rather than scopes. If pass B ran on targets the user had already ruled out on country, funding, or deadline, the search budget was spent on nothing.

Volatility

Themes, specialisation names, pathway lists, and topic restrictions change every cycle, and institutions differ within every country. This skill encodes the extraction method only — there is deliberately no catalogue of programmes and their themes, because such a catalogue would be wrong within a cycle. Re-extract per cycle.