Claude Scholar is a semi-automated research assistant distributed as a Claude Code plugin pack for academic research and software development, primarily targeting CS/AI researchers. It orchestrates the full research lifecycle (ideation, literature review, experiment analysis, paper writing, reviewer rebuttal, post-acceptance promotion) through 40+ markdown skills, slash commands, agents, lifecycle hooks, and Python/Node.js scripts. Despite the academic domain, this is an application repo: the deliverable is an installable plugin (scripts/setup.sh) with a manifest-based uninstaller, GitHub Actions CI, and executable code, not a data pipeline or notebook collection. Architecturally it is a plugin bundle, not a microservice or monolith. Entry points include scripts/setup.sh (installer), hooks/hooks.json (Node.js lifecycle hooks: security-guard, skill-forced-eval, session-start/session-summary/stop-summary), commands/ (slash commands such as /kb-init, /kb-sync, /analyze-results, /rebuttal), agents/ (literature-reviewer, paper-miner, rebuttal-writer, code-reviewer), and skills/ (self-contained packages each with SKILL.md, references/, examples/, scripts/). The most engineered subsystem is the Obsidian knowledge base: a Python CLI (project_kb.py plus kb_common.py) manages a vault-first project lifecycle with a machine-parseable registry, auto-index regeneration, note templates, and link linting. External dependencies include the Zotero MCP server for literature import, arXiv/CrossRef/Semantic Scholar APIs for citation verification, and an Obsidian vault for durable knowledge storage. Primary data flows are human-in-the-loop research pipelines: (1) Zotero to Sources/Papers to Knowledge to Writing to literature.canvas, with an enforced Evidence Record contract (claim strength: speculative/observed/supported/strong) gating what may be written; (2) experiment analysis producing strict report bundles (analysis-report, stats-appendix, figure-catalog) with mandatory statistical rigor; (3) citation verification running a four-layer match against scholarly APIs. The repo also contains a meta-layer for building Claude Code plugins (skill-development, command-development, hook-development, skill-quality-reviewer), forming a self-improvement loop. Maintenance is active with recent commits and CI, but technical drift is medium: multilingual docs (EN/ZH/JA) risk divergence, venue templates and competition analyses are future-dated and will age, machine-specific paths are committed, some code paths are marked reference-only, and external MCP/API dependencies can break. Overall this is a well-structured and unusually comprehensive plugin collection with several genuinely reusable components, though much of the corpus is markdown-as-code rather than executable software.