Academic research

Research a specific question or topic at university level. Output must be facts, quotes, and sources — not unsourced summaries or invented citations.

Core rules

1. Every claim needs a source. Prefer primary papers over secondary blogs.
2. Bind claims to evidence. Quote or closely paraphrase a specific passage;

note page/section when available. No orphan assertions.

1. Prefer originals on conflict. If two sources disagree, re-check the

primary document (or PDF) before resolving; say when unresolved.

1. Separate fact / interpretation / gap. Label clearly what the source

states vs. what you infer vs. what is missing.

1. Never invent citations. If you cannot retrieve or open a source, mark it

unverified and do not cite it as evidence.

1. Never treat training memory as retrieved evidence. Only cite what you

fetched in this session (API, catalog, or opened full text).

Budgets (default)

Stop expanding search unless the user asks for more:

Agent tools

1. Discovery / metadata: Shell + curl (or equivalent HTTP) against the

APIs in [references/apis.md](references/apis.md). Prefer curl over guessing.

1. HTML landing pages / abstracts: WebFetch when the API already gave a

stable URL.

1. PDFs: download with curl to a temp file, then extract text (below). Do

not claim page quotes from metadata alone.

1. Keys: read from env if set — never invent keys; never commit them.

If a required key is missing and the unauthenticated path fails, say so and continue with the next API in the fallback order.

Fallback order

1. OpenAlex — default discovery when no S2 key (search often 429s on the

shared pool).

1. Semantic Scholar — citation graph + recommendations; use key if

available. On 429: sleep ~2s once; if still 429, switch to OpenAlex for search; keep S2 for single-paper / citations only if those still succeed.

1. Crossref — DOI resolution / bibliographic confirmation.
2. BASE — only with BASEAPIKEY; German/EU repositories & theses.
3. DNB — German national bibliography / authority (no key).

Do not block the whole research on one failing API.

Conceptual workflow

Copy and track:

Research progress:
- [ ] 1. Frame the question
- [ ] 2. Multi-perspective sub-questions
- [ ] 3. Discover candidate sources (APIs / catalogs)
- [ ] 4. Triage by relevance + impact
- [ ] 5. Read / extract evidence (quotes, numbers, methods)
- [ ] 6. Citation-graph pass (pioneers vs follow-ups)
- [ ] 7. Adjudicate conflicts
- [ ] 8. Synthesize answer with sources

1. Frame the question

• Restate the user's question as a clear research question (empirical questions

should be answerable with evidence; conceptual/historical ones with scoped claims).

• Define scope: field, geography/language (e.g. DE vs EN), time window, OA-only.
• List inclusion/exclusion criteria before searching.

2. Multi-perspective sub-questions

Generate 3–6 angles before deep reading, e.g.:

• definition / state of the art
• methods & measurement
• empirical findings
• critiques / limitations
• policy or practice implications (if relevant)

Search and answer per sub-question, then merge.

3. Discover candidate sources

Use scholarly APIs first (not general web search alone). Details: [references/apis.md](references/apis.md).

Typical pass: OpenAlex search → hydrate top hits → optional S2 citation graph on the best 2–3 → Crossref if DOI metadata is thin → BASE/DNB only if DE catalogs matter.

4. Triage by relevance + impact

Keep 8–12 candidates; deep-read 3–6.

5. Extract evidence (quotes)

1. Capture abstract-level claims from API fields first; mark location abstract.
2. Resolve a PDF URL: S2 openAccessPdf.url, or OpenAlex

bestoalocation.pdfurl / openaccess.oa_url.

1. If no PDF: quote only from abstract/API text, or mark claim as

metadata-only (weaker).

1. If PDF available:

curl -fsSL -o /tmp/paper.pdf 'PDF_URL'
# Prefer pdftotext when installed; else python pypdf/pdfplumber if available
pdftotext -layout /tmp/paper.pdf - | head -n 200

For complex multi-column layouts, see [references/frameworks.md](references/frameworks.md) (Docling). Do not invent page numbers — use abstract or § headings if pages are unknown.

Per kept source, fill:

Markdown-first: put n, p, CI, effect sizes in tables.

6. Citation-graph pass

For the top 2–3 papers (Semantic Scholar):

• /references → foundations (parse citedPaper)
• /citations → extensions / critiques (parse citingPaper)

Distinguish pioneer vs follow-up before treating a paper as definitive.

7. Adjudication

When sources conflict:

1. Prefer the primary study over reviews summarizing it.
2. Prefer fuller methods reporting and larger / more appropriate samples.
3. Re-open the original PDF when metadata or secondary summaries disagree.
4. If still unclear: report both sides with citations; do not pick a winner.

8. Methodology critic (when empirical)

Check Methods for: design, sample size, significance / uncertainty, confounds, generalizability. Flag weak evidence explicitly.

Output format

# [Research question]

## Short answer

2–5 sentences. Every factual sentence either cites a source or is clearly
labeled as synthesis.

## Findings

### [Sub-question / theme]

- **Claim:** …
  - **Evidence:** "…" (Author et al., Year, p./§…)
  - **Source:** DOI / URL / catalog ID
  - **Caveat:** … (optional)

## Source list

| #   | Citation | Year | Type | DOI / ID | OA PDF | Why included |
| --- | -------- | ---- | ---- | -------- | ------ | ------------ |

## Gaps and open questions

What the literature does not settle.

## Search log (brief)

APIs, queries, filters, HTTP failures (e.g. 429), date of search — enough to reproduce.

Tooling priorities

1. Metadata + discovery: OpenAlex, Semantic Scholar, Crossref; BASE, DNB (DE).
2. Full text: OA PDF URLs from APIs → extract text → quote.
3. Optional deeper tooling: [references/frameworks.md](references/frameworks.md).

API details → [references/apis.md](references/apis.md)