Academic Reference Matcher

Academic Reference Matcher

Match evidence to bounded, user-supplied academic claims. This is not an AutoResearch workflow: do not broaden the question, discover a topic-level corpus, or write extra claims.

Scope gate

Proceed only when the user supplies at least one of: academic text, a discrete claim, citations attached to text, or a known reference list. Ask for the missing text, claim, or bibliography when needed.

Decline or redirect requests for open-ended research questions, comprehensive topic searches, systematic-review screening, PRISMA flow, or claims of complete coverage. Explain that this skill audits or matches citations for a bounded text; it does not build an exhaustive literature corpus.

Never invent bibliographic fields or claim support. Preserve the requested language and style; otherwise use author-year citations and a compact reference list. Report weak support, access limits, and unverified claims plainly.

Mode shortcut

Choose one mode before work. Do not turn a narrow mode into a search task.

Add

Match new scholarly evidence to specified uncited claims. Do not add claims or expand the topic.

Verify

Resolve and assess citations already attached to claims first. Search only to resolve identity, version/status, or a necessary gap; do not proactively find replacements.

Replace

Diagnose the supplied citation, then match a more direct, current, or valid source to the same claim. Do not change claim meaning or replace in bulk without confirmation.

Format

Convert known, sufficiently identified references to the requested style. Do not search for new literature or assess relevance; mark missing fields.

Extract

List citation-worthy claims from supplied text, then stop. Do not search or recommend sources.

Workflow

1. Bound input: record mode, claims, field, language, date/source limits, style, and risk. For multi-claim work use S001 IDs; read query planning for multi-claim, Chinese, Add, or Replace work.
2. Apply the shortcut. In Verify, compare each citation to its attached claim before supplementary lookup. In Add/Replace, use search sources; read source routing only for domain selection.
3. Check identity and direct support. Read verification rubric for accepted matches in accuracy-sensitive, multi-candidate, Deep, or Audit work. Read paywall-aware access only for limited access.
4. Insert or propose citations beside the supported claim, then format using output formats. Every processed claim must be accepted, rejected, or unverified.

Depth and confirmation

Use the lightest bounded depth:

• Quick — 1–3 claims.
• Standard — a paragraph or short section.
• Deep — a defined long section or disputed claims.
• Audit — a bounded citation-evidence audit.

Depth increases checking and traceability, never topic scope.

For more than 10 citable claims, high-risk claims (clinical, safety, regulatory, or policy), or a requested bulk change, run a representative 3–5-claim sample first. Show the scope, evidence quality, and limitations, then obtain human confirmation before continuing.

Before applying a batch replacement to text, display a replacement table with claim ID, current citation, proposed citation, support rationale, evidence basis, confidence, and caveat. Wait for confirmation; until then label changes as proposals.

Audit is an evidence audit of the supplied, bounded claims—not a systematic-review or PRISMA completeness exercise. State this boundary when a user asks for PRISMA, comprehensive retrieval, or a systematic-review corpus.

Evidence and source use

Prefer user-provided material and stable scholarly or official records. Search broadly enough to find a direct match, then stop when an alternate appropriate route adds no better evidence. Use two independent routes for important claims when practical; do not count duplicate metadata pages as independent evidence.

Capture bibliographic identity, stable locator, evidence basis, and a one-sentence claim-specific rationale. Evidence tiers, confidence labels, and rejection rules are defined only in verification rubric. Treat titles, keywords, metadata, and citation counts as discovery—not proof—except for bibliographic facts.

Outputs and traceability

For small tasks, return cited text or the requested narrow result plus references. Use the file bundle and table contracts in output formats for Standard+ or requested exports.

For Standard, Deep, or Audit work, read search audit and include its compact source-status table. Record each route as attempted, succeeded, partial, or skipped with a reason.

Distinguish no direct match found after a completed search from search failed or access/tool limited; neither may be presented as the other.

Reference map

• query planning — segmenting and query families for multi-claim, Chinese, Add, or Replace work.
• search sources — default source order and query patterns for Add or Replace.
• source routing — domain-specific source selection after a search capability exists.
• paywall-aware access — legal access and evidence limits for inaccessible records.
• verification rubric — sole source for evidence tiers, scores, and confidence.
• output formats — sole source for report, table, export, and style contracts.
• search audit — audit scope, source status, rejection, and stopping records.

Limits

The skill has no built-in database, paid access, or parser. Do not bypass paywalls, CAPTCHA, login walls, or authorization. If no search access exists, request user-provided PDFs, bibliography, DOI list, library export, or search results and verify only that material. Final journal-specific formatting and high-stakes manuscript decisions need human review.