research

Use whenever the user is doing research, benchmarks, or comparisons that involve TWO OR MORE AI models, agents, or LLMs (closed or open source). Enforces experimental parity across model tier, sampling configuration (temperature, max_tokens, reasoning effort, system prompt, seed), and cost transparency. BLOCKS execution until parity is verified or the user has explicitly approved an asymmetry. Activates on phrases like "compare GPT-4 vs Claude", "benchmark these models", "evaluate Llama against Gemini", "which model is better at X", or any research workflow that produces side-by-side AI outputs.

Use whenever the user is writing, reviewing, or editing research that contains citations to academic papers, preprints, books, or technical reports. Verifies (1) the cited work actually exists, (2) the metadata (authors, title, year, venue) is accurate, and (3) every claim attributed to a cited work is genuinely supported by that work. BLOCKS inclusion of any unverified citation. Activates on phrases like "add a citation for", "write the related work section", "check my references", "is this paper real", "did this paper actually say", "review my bibliography", or any task producing or modifying a reference list.

Use when the user wants a literature search across multiple sources for a research question, NOT for summarizing a single known paper (that's paper-summarizer). Performs multi-pass search across arxiv, OpenAlex, and Semantic Scholar, deduplicates results, runs every candidate through citation-check before it can be cited, and optionally produces structured summaries via paper-summarizer. Activates on phrases like "find papers on", "what's the literature on", "do a lit review for", "find recent work in", "search for papers about", or any task that produces a multi-paper bibliography.

Use whenever the user wants a structured summary of a specific academic paper (arxiv preprint, DOI-resolved publication, technical report, or local PDF). Produces a fixed-schema summary covering claim, evidence, methodology, limitations, and replication effort, with EVERY extracted point grounded line-by-line in the source paper. Activates on phrases like "summarize this paper", "what does <paper> say about", "TLDR this arxiv link", "give me the key claims from", or any URL/DOI/arxiv ID handed in with intent to digest. NOT for literature search across multiple papers — that's lit-review.

Use when Sidd is about to submit, share, or publish a research artifact (paper, report, blog post, thesis chapter, slides) and wants a final integrity sweep. Orchestrates citation-check, replication-pack, and a structured manuscript audit, blocking on any unresolved issue. Activates on phrases like "ready to submit", "final version", "before I submit", "pre-submission check", "is this ready", "submission deadline", or any signal that the user is finishing a research artifact. Should also be PROACTIVELY OFFERED whenever the user appears to be wrapping up a research project — ask "want to run a pre-submission check?" before they proceed.