WenyuChiou/ai-research-skills

What it solves

AI research assistance often fails in five concrete ways, and this catalog is built to fix them specifically. First, every new chat session makes you re-explain your research context — the project stores that state in checked-in manifest files instead. Second, AI models hallucinate citations that don't exist; this project forces any claim without supporting evidence artifacts to be labeled as a gap, not a fact. Third, reading dozens of papers still doesn't tell you which research gap is worth a thesis — the gap-to-topic skill applies three structured gates (is it open? is it a real contribution? is it feasible?). Fourth, AI-written prose is easy to spot and harms manuscript credibility; dedicated writing and evidence-audit skills address that. Fifth, switching between AI tools like Claude, Codex, and Antigravity wipes your state — explicit, durable handoffs solve that too.

How it works

The core asset is a set of 17 portable skill files, each defining a specific research task as a SKILL.md contract. These skills are organized around an eight-stage pipeline: discover literature, find the gap, frame the research question, plan the project, build the model, run and validate, draft the manuscript, and submit plus respond. Each stage's output becomes the next stage's input — the handoff is mechanical, defined by YAML schemas rather than free-form text. A downstream skill can refuse to process malformed input, which prevents overconfident claims from propagating. Three design principles carry the whole system: manifests store research state in check-in YAML/Markdown files so new AI sessions re-onboard themselves; schemas with anti-leakage rules structurally separate supported claims from gaps; and character-driven routing delegates token-heavy mechanical work to Codex or Antigravity while keeping judgment, long-context synthesis, and final review with the primary model. A separate research-hub runtime (installable via pip) provides an executable, resumable workflow, and an optional agent-collab harness adds policy and checkpoint governance. Human-in-the-loop controls pause automation at five gates: breaking contracts, skill removal, external writes, semantic acceptance, and release authorization.

Who it's for

This is built for graduate students, PhDs, postdocs, research support staff, and agent/harness builders who want inspectable research workflows rather than a black box. It's also relevant for librarians, research assistants, or advisors helping others adopt AI for research — they can share the README without installing anything. Agent and harness builders get a dedicated architecture guide covering state, human-in-the-loop controls, recovery, and evaluation.

Highlights

• 17 portable skills installable as plugins, starting with a 12-skill research workspace and adding writing, Zotero, or delegation plugins only when needed
• Installable via one command: claude plugin marketplace add WenyuChiou/ai-research-skills, then claude plugin install research-workspace@ai-research-skills
• Uses plain-language requests: describe what you want (e.g., "Compare these 5 papers by method, data, limitations") and the right skill activates
• Integrates Zotero, Obsidian, and NotebookLM as one optional library-management loop — all three are optional and the system adapts to whichever subset you have
• Works across many hosts: Claude Code, Codex CLI, Cursor, Windsurf, Gemini CLI, Hermes, and OpenClaw
• Includes a live CLI workflow: pip install research-hub-pipeline, then research-hub setup and research-hub doctor
• Anti-hallucination enforcement: claims with empty evidence artifacts are forced to carry status: gap plus a gap reason, never supported
• Human-in-the-loop gates ensure read-only and reversible steps continue automatically while breaking contracts and external writes require explicit human approval