Yucheng Zhang

Compress the current session into a HANDOFF.md in the working directory so the next agent can resume without loss of context. Run this at the end of any significant session.

Structured hypothesis formulation from observations. Use when you have experimental observations or data and need to formulate testable hypotheses with predictions, propose mechanisms, and design experiments to test them. Follows scientific method framework. For open-ended ideation use scientific-brainstorming; for automated LLM-driven hypothesis testing on datasets use hypogenic.

Detect methodological collision between a draft paper and recent literature on arXiv / Semantic Scholar / OpenAlex. Extracts the draft's methodological features, runs iterative semantic search via paper-lookup, scores each candidate on 4 overlap axes, and produces a ranked collision-risk report with defense suggestions. Use when the user says "check for methodological overlap", "collision detection", "pre-submission competitive analysis", or invokes /lit-collision. NOT for full literature review (use literature-review) or generic paper search (use paper-lookup).

Conduct comprehensive, systematic literature reviews following PRISMA methodology. Use when the user asks for a systematic review, meta-analysis, scoping review, research synthesis, or the literature-review section of a paper/thesis. Produces a structured Markdown document with documented search strategy, PRISMA flow counts, thematic synthesis, quality assessment, and verified citations.

Run an adversarial multi-model review on a proof, non-trivial diff, or architecture decision: primary verifier → adversarial stress-tester (different model family) → primary arbitrator. ACTIVATE ONLY when the user explicitly asks for 'rigorous review', 'adversarial review', 'stress test', 'multi-model review', 'check this proof', or names a specific high-stakes artifact (migration, security fix, proof, algorithm) that must not ship with a silent bug. Do NOT use for: trivial edits, formatting changes, docs-only PRs, exploratory prototypes, general 'can you take a look' requests, or cases where a single-model self-review is clearly sufficient. Works with any external-model CLI that can attach files or a git diff.

Look up a specific paper or exact-known query across 10 academic databases via their REST APIs (PubMed, PMC, bioRxiv, medRxiv, arXiv, OpenAlex, Crossref, Semantic Scholar, CORE, Unpaywall). Trigger ONLY when the user supplies a precise identifier (DOI, PMID, PMCID, arXiv ID, exact title) OR explicitly names one of these databases. Use for: fetching metadata for one known paper, pulling an abstract or full text for a known article, checking open-access status, retrieving a citation count, listing an author's publications, fetching references or citing papers for one specific paper. Single-shot targeted queries only — do NOT trigger on broad topic discovery ("find papers about X"), systematic review workflows, or methodological overlap scanning; those are handled elsewhere.

Audit the theoretical appendix of a math/statistics paper or thesis for logical gaps, unjustified steps, missing assumptions, and edge cases. Output a structured fix plan with copy-pasteable LaTeX patches anchored by gap IDs. Use when the user says "audit this proof", "check appendix rigor", "find gaps in the proofs", or invokes /proof-audit. Best for pre-submission thesis review. NOT for teaching proofs (use explain-paper) or full-paper critique (use paper-reader).

Maintain a persistent, Karpathy-style research wiki per domain via three operations: ingest (compile a paper into the wiki), query (answer a question against the wiki), lint (audit pages). ACTIVATE when the user says 'ingest this paper', 'add to wiki', 'query the wiki', 'lint the wiki', 'what does my wiki say about X', or when working inside a directory that already contains a `raw/` + `wiki/` + `CLAUDE.md` layout. Do NOT use for: general literature search (use a paper-lookup skill); one-off paper reading where persistence is not wanted; any directory that does NOT already have the `raw/` + `wiki/` + `CLAUDE.md` layout (bootstrap a new wiki via `scripts/new-wiki.sh` first). Not a RAG system — the wiki is LLM-compiled Markdown, browsable in Obsidian.

Review the current git diff for reuse, dead code, over-abstraction, premature optimization, and needless complexity, then apply fixes in place. ACTIVATE when the user says 'simplify', '/simplify', 'clean this up', 'polish', 'review your own work', or 'tighten this diff'. Do NOT use for: docs-only changes, pure renames, whole-codebase refactors (this skill is scoped to the current diff only), or when no code has changed yet. Scans `git diff` (default: HEAD) — if there is no diff, exits without action.

Fetch web pages that built-in WebFetch / WebSearch cannot read — Cloudflare or PerimeterX walls, JS-rendered SPAs, paywalled or login-gated content, sites that block by geo/TLS fingerprint, 403/429/503 endpoints. Use whenever WebFetch returns an error, empty body, CAPTCHA / challenge stub, 'please enable JavaScript' page, or obviously-truncated content. Primary entry point: scripts/fetch.sh <url>, which runs a Jina Reader → stealth curl → Wayback Machine fallback chain biased by per-domain hints from references/site-patterns/. For multi-URL research fanout, use scripts/fetch-parallel.sh which runs N independent fetches concurrently and emits a manifest. If the URL points to a PDF, the script exits with code 2 and defers to your PDF-extraction tool. For login-gated or interaction-heavy pages, escalate to the chrome-devtools MCP tools.