Academic Citation Workflow

Citation work has two dominant failure modes: hallucinated or mischaracterized references, and silent budget blowups (page count, table scope, row sprawl). Both are prevented structurally, not by care alone. This skill encodes a workflow validated on a real IEEE journal submission (one screening round cutting 17 weak references, one enrichment round admitting 8 web-verified references; the verification layer caught fabrication risks, metadata errors, and one factual error in the paper's own table).

Fix admission criteria before any search

Write the criteria into the search prompts themselves, before launching: domain scope (e.g., "strictly LLM/AI-traffic systems"), a notability floor (venue tier for academic work, deployment scale for industrial), year preferences, and, for comparison tables, the rule that a candidate is admissible only if it takes a real value in every comparison column. Mid-flight scope corrections from the user mean a wasted round; ask once up front.

Finding references (enrichment)

Run a three-stage pipeline: per-category research agents, then one adversarial verifier per candidate, then synthesis.

• Research agents get one category each and a hard rule: a candidate may only be returned together with a URL the agent actually fetched and read. This single constraint removes most hallucination risk at the source. Prefer recent top-venue papers and widely deployed industrial systems with official documentation.
• Verifiers re-search from scratch and do not trust the researcher's URL. Each verifier checks four things: the work exists with the claimed title/venue/year; it actually does what the summary claims; it genuinely fits the claim or table row it would support; the BibTeX metadata is correct (authors, venue, year). Have verifiers also audit the surrounding existing content while they are there, since independent verifiers catch the paper's own pre-existing errors (a mischaracterized system, a wrong table cell), which is free review.
• Cite industrial systems as @misc with the official docs URL, a double-braced corporate author, and an accessed-year note. Their citation value is proving that a design class is deployed in practice, so the official page beats a third-party writeup.

Minimal Workflow-tool skeleton for the pipeline: