WenyuChiou

Use research-hub to operate Zotero, Obsidian, and NotebookLM research workflows through CLI, MCP, REST, and dashboard. Trigger when the user asks to find papers, ingest sources, organize a literature review, build Obsidian research notes, upload sources to NotebookLM, inspect clusters, generate AI briefs, or maintain a research vault.

Turn a list of papers (Zotero collection, Obsidian cluster, manual list) into a compact comparison matrix written to .research/literature_matrix.md, instead of generic per-paper summaries. Use when the user asks to "make a literature matrix", "compare these papers by method/data/limitations", or "decide which papers are central to my review". If the user says "extract the claims from these papers": cross-paper comparison matrix → this skill; claims from their own manuscript draft → `paper-memory-builder`; per-cited-paper Key Findings → `paper-summarize`. This skill compares a KNOWN paper set; to turn a research area into a go/no-go decision on a candidate thesis/proposal topic, use `gap-to-topic`.

After research-hub ingests a cluster of cited papers, fill the per-paper Key Findings + Methodology + Relevance sections in BOTH Obsidian markdown and the Zotero child note. Use when the user says "fill the TODO Key Findings/Methodology blocks left by research-hub auto", "I just ran auto and don't know what these papers are about", or "summarize the papers in cluster X". Invokes a supported LLM CLI on each paper's abstract. NOT for summarizing the user's own manuscript draft — that's `paper-memory-builder`. NOT for cluster-level briefs — that's `research-hub notebooklm generate`. This skill is per-cited-paper only. If the user says "extract claims from these papers", disambiguate before acting: their own manuscript draft → `paper-memory-builder`; a cross-paper comparison matrix → `literature-triage-matrix`; per-cited-paper Key Findings → this skill.

Operate research-hub workflows for literature discovery, source ingest into Zotero/Obsidian/NotebookLM, dashboard inspection, and vault maintenance. Use when the user asks to find papers and organize them, build a knowledge base, ingest a folder of PDFs, upload to NotebookLM, generate research briefs, inspect clusters, or maintain a research vault. NOT for auditing or cleaning up an existing Zotero library — that's `zotero-library-curator` (read-only audit) plus `zotero-skills` (for CRUD).

Research-domain router that writes `.coord/multi_ai_plan.md` when a single round of work will need two or more delegates AND the work touches research-hub artifacts (`.research/`, `.paper/`, Zotero/Obsidian/NotebookLM pipelines). For a single delegate, use `codex-delegate` or `gemini-delegate` directly — do not invoke this skill. For generic non-research multi-agent decomposition (pure code refactor, generic translation, no research-hub artifact), use `agent-collab-workspace:agent-task-splitter` instead (different artifact `.coord/plan.yml`, different scope). The router decides task splitting, dependency ordering, and reconciliation; the leaves execute.

Audit and curate a Zotero library — find duplicate DOIs, orphan items missing required tags, propose collection rebinds, identify bloated or under-used collections, generate tag hygiene reports, emit preview-only cleanup plans. Use when the user asks to "audit Zotero", "find duplicates", "tag hygiene report", "which collections are bloated or under-used", or "propose a Zotero cleanup plan". Defers all CRUD operations to the standalone `zotero-skills` skill or `research-hub zotero` CLI. Includes a backup-first reminder before any apply/CRUD handoff suggestion.

Convert a paper draft + figures + Zotero metadata into reusable .paper/claims.yml and .paper/figures.yml files so the academic-writing-skills skill can do writing, revision, and audit passes without re-reading the manuscript every time. Use when the user asks to "build paper memory", "extract claims from this manuscript", "extract claims, supporting evidence, and figure key numbers", or "prepare this paper for AI-assisted writing". NOT for summarizing cited papers in a literature cluster — that's `paper-summarize`. This skill is for the user's own manuscript draft only.

Guide a researcher through 5 Socratic segments — research question sharpening, expected mechanism, identifiability check, validation plan, risk register — and produce `.research/design_brief.md`. Use when the user asks to "frame this research question", "design my study", "help me think through what model to build", "sharpen my hypothesis", "is my research question sharp enough to be falsifiable?", or "before I start coding, walk me through the design". Runs AFTER a topic is chosen — it designs the study for a given question; to decide whether a research gap is worth pursuing at all (open / a contribution / feasible), use `gap-to-topic`. Does NOT write the model spec; does NOT invent the research question — guides the human to articulate them.

Use when the user asks to split a goal across Claude, Codex, or Gemini (gemini requests are rerouted - lane deprecated); plan a multi-agent run; break work into parallel agent tasks; or decompose a large task that needs bounded context handoffs. This is the **generic** multi-agent task splitter — writes `.coord/plan.yml` (a DAG) plus per-agent task files. NOT for research-domain routing that touches `.research/`, `.paper/`, or Zotero/Obsidian/NotebookLM ingest pipelines — for those, use `research-hub-multi-ai` instead (different artifact `.coord/multi_ai_plan.md`, research-hub-aware reconciliation).

Use when a multi-agent round needs a pre-merge gate, pre-commit check, verification before push, or a PASS/FAIL decision after reconciliation.

Use when multiple agents have completed a round and the user asks to reconcile outputs, compare Codex and Gemini, synthesize run results, identify conflicts, or decide what should be retried.

Use when multi-agent work risks context overflow, memory growth, noisy logs, oversized handoffs, cross-session continuation, or parallel Codex and Gemini execution.

Use when a task needs single-agent self-correction across multiple iterations — write plan, execute, critique own output, revise plan, re-execute, until convergence or budget exhausted. Different from `agent-debate` (which is 2 agents arguing pro vs con); this is 1 agent looping over its own work.

Use when a consequential decision needs adversarial review, opposing agent arguments, a second opinion via debate, or explicit trade-off analysis before implementation.

Use when the user asks to update shared memory, initialize multi-agent memory, summarize decisions so far, identify open questions, or prepare a fresh session primer.

Operating harness for working on the method-harness-compiler project (or reviewing its artifacts). Use when starting ANY task on that project — phase reviews, tool discovery, PR review, eval design, memory updates, runtime export, repo maintenance, or A/B test design. Routes the agent to a minimal per-task file set instead of whole-repo reading. For work on ANY OTHER project, activate only when the task is long / multi-step, multi-agent, expensive-if-wrong, makes a completion claim (done / passing / fixed / ready / safe / staged), or touches governance (permissions, hooks, cron, CI, routing) — the portable entry is then core/GLOBAL_BOOTSTRAP.md. Skip for one-line / typo / format edits and pure capability asks.

Use this skill when auditing, simplifying, benchmarking, or refactoring an AI/agent project harness, including CLAUDE.md, AGENTS.md, slash commands, skills, subagents, hooks, tool routing, scheduled reviews, AI-review reports, benchmark scaffolds, and rolling harness maintenance. Not for ordinary code review of application code (use code-review) or for running the method-harness-compiler project (use the root harness ladder).

When the user is building a tool-calling agent and gets stuck — "為什麼 LLM 不呼叫我的 tool", "我這 schema 哪裡寫壞", "tool 被呼叫但 args 不對", "ReAct loop 跑不停", "the LLM won't call my tool", "help me design a function schema", "debug this tool-use behavior". Walks them through a 4-branch diagnostic + 5-step schema design walkthrough, with references to bad/good schema A/B and SDK-diff cheatsheet. Do NOT use for: pure LangChain / LangGraph / CrewAI framework questions (route to Stage 4 frameworks), MCP server building (route to cookbook 2), production agent observability (route to Stage 7).

Delegates implementation-heavy or repetitive coding work (batch edits, boilerplate, multi-file refactors with clear patterns, test scaffolding) from Claude to OpenAI Codex CLI. Use when token cost outweighs judgment cost. Trigger phrases include "delegate to codex", "let codex do this", "batch refactor across files", "scaffold tests for". Avoid for architecture, security review, or root-cause debugging.

Full CRUD operations on Zotero library — search, add, update, delete items with notes, tags, and collections (PDF upload is available via raw pyzotero, see references/create-operations.md). Uses dual-API architecture (local API for fast reads, Web API for writes). Use this skill whenever the user mentions Zotero, references, citations, literature management, reading notes, or wants to organize academic papers — even if they don't explicitly say 'Zotero'.

Delegates large-context reading, bilingual or Chinese (CJK / zh-TW) drafting, cross-file synthesis, and second-opinion review to Google Antigravity CLI (`agy`) or legacy Gemini CLI. Use when input exceeds Claude's working budget, when the user writes in Chinese, when terminology must align across long documents, or when a reviewer pass is needed. Trigger phrases include "summarize this in Chinese", "second-opinion review", "long-context synthesis", "draft this in zh-TW". Avoid for bulk code generation or security-sensitive coding.

End-to-end academic manuscript workflow for drafting, revision, reviewer response, figure-text consistency, claim-evidence audits, and pre-submission checks. Use this skill whenever a user asks for manuscript sections, paper revision, rebuttal letters, journal compliance, overclaim detection, GPT-style prose cleanup, figure captions, or evidence-backed academic writing. It is especially useful for multi-section papers where context, claims, figures, and reviewer comments must stay consistent across revisions. When `.paper/claims.yml` and `.paper/figures.yml` exist (produced by the `paper-memory-builder` skill), use them as ground truth instead of re-reading the manuscript; if they are absent and the task is multi-section / claim-evidence / reviewer-response work, suggest running `paper-memory-builder` first to build the memory layer.

Sweep up ALL session artifacts left behind by Claude Code: stale git worktrees, dangling git refs, and orphaned worktree directories. Use this skill whenever the user mentions slow performance, low disk space, too many worktrees, cleanup, maintenance, freeing space, or wants to reduce clutter from Claude Code sessions. Also triggers for 'my computer is slow', 'running out of space', 'clean up after Claude', or any mention of worktree or session accumulation. Especially relevant for heavy Claude Code users who spawn many code tasks — each task creates a full git worktree copy that persists after the session ends.

Delegate lane for Google Antigravity CLI (agy) - bounded, non-honesty-critical mechanical subtasks via a brief file, an alternative cheap executor alongside codex-delegate and claude-cheap. Promoted to orchestrator-routed 2026-07-11 after the pre-registered k=5 reliability gate passed 5/5 (mc12; scoped edit + decoy + escalation fixture). Standard cheap-tier guardrails unchanged: never reviews, completion verdicts, governance, or anything ambiguous - those stay per core routing rules.