ARS (Academic Research Skills) is a skill package for Claude Code positioning itself not as a prompt collection but as a multi-agent orchestration framework for academic work: 42 agents across 4 skills and 25+ modes, connected by a pipeline scheduler covering research → writing → review → revision → publication. Its core assumption is that human-AI collaboration beats full automation, making it a quality-control system rather than an auto paper generator.
Architecture
- Agent system: 42 specialized agents with defined responsibilities and constraints
- Mode system: 25+ operation modes (full / socratic / revision / calibration, etc.)
- Scheduler: a 10-stage pipeline with checkpoints and quality gates
- Schema layer: handoff schemas defining cross-agent data formats (Material Passport)
- CI/Lint: GitHub Actions-driven spec consistency checks
- External verification: Semantic Scholar API, DOI validation, VLM figure checks
- 7-category failure-mode checklist (v3.2): blocking checks at both integrity gates — the pipeline halts until the user confirms any suspected failure
- 5-type citation hallucination taxonomy (v2.7): Title Fabrication, Partial Author Claim, Inaccurate Hallucination, Partial Hallucination, Synthesis Hallucination
- Devil's advocate concession threshold (v3.0): rebuttals must score ≥4/5 before the DA concedes; no consecutive concessions; frame-lock detection
- Anti-sycophancy (v3.0/v3.5.0): conversational health self-checks every 5 turns; a Collaboration Depth Observer scoring 4 dimensions, requiring concrete dialogue-turn citations for scores ≥7
- Cross-model verification (v3.0):
ARS_CROSS_MODELenables 30% sampled independent review by a second model (GPT-5.4 Pro or Gemini 3.1 Pro) - Material Passport: YAML audit trail recording
literature_corpus[], append-onlycompliance_history[],reset_boundary[]for cross-session recovery, optionalrepro_lock - Intent-based socratic routing: mode routing detects user intent signals, not keywords — works in any language
- SessionStart announce hook: full context injection on
startup/clear, one-line confirmation onresume/compact— precise context burn control - GitHub: https://github.com/Imbad0202/academic-research-skills
- Latest version: v3.7.0 (2026-05-05); Author: Cheng-I Wu; License: CC-BY-NC 4.0
- Codex CLI sibling: https://github.com/Imbad0202/academic-research-skills-codex
- Lu, C. et al. (2026). "Towards end-to-end automation of AI research." *Nature* 651, 914-919.
- Zhao, X. et al. (2026-05). Citation hallucination audit. arXiv:2605.07723.
- Song, Y. et al. (2026). "PaperOrchestra." arXiv:2604.05018.
- Wang, S. & Zhang, H. (2026). "Pedagogical partnerships with generative AI in higher education." *IJETHE* 23:11.
| Skill | Agents | Responsibility | Version | |-------|--------|----------------|---------| | Deep Research | 13 | Literature review, PRISMA systematic reviews, socratic guidance | v2.8 | | Academic Paper | 12 | Paper writing, style calibration, LaTeX output | v3.0 | | Academic Paper Reviewer | 7 | Multi-perspective peer review, 0-100 scale, devil's advocate | v1.8 | | Academic Pipeline | 10 | Full-process orchestration, integrity gates | v3.7 |
Pipeline stages include drafting, a mandatory Stage 2.5 integrity gate, peer review, re-review, revision, a final integrity gate (Stage 4.5), and finalization with process logging and AI self-reflection.
Why Human-AI Collaboration, Not Full Automation
ARS diverges fundamentally from The AI Scientist (Lu et al., 2026, *Nature*), the first end-to-end automated AI research system whose paper passed ICLR 2025 workshop blind review. Lu et al.'s own limitations section lists 7 structural failure modes: implementation bugs passing AI self-review, hallucinated experimental results, shortcut reliance, bug-as-insight reframing, methodology fabrication, frame-lock, and citation hallucination. ARS bets that a human researcher + AI combination avoids these better than either alone.
Corpus-scale evidence supports the concern: Zhao et al. (2026-05) audited 111 million citations across 2.5 million papers (arXiv, bioRxiv, SSRN, PMC), conservatively estimating 146,932 hallucinated citations in 2025 alone, with an 85.3% survival rate of hallucinations through the bioRxiv-to-PMC published pipeline. This directly motivated ARS v3.7.x's three-layer citation anchoring.
ARS v3.3 also adopted four PaperOrchestra (Song et al., 2026, arXiv:2604.05018) techniques: Semantic Scholar API validation (Levenshtein ≥ 0.70 title matching, DOI mismatch detection), anti-leakage protocol (marking [MATERIAL GAP] instead of filling from LLM memory), VLM figure verification with a 10-item checklist, and per-dimension score trajectory tracking across revision rounds.
Key Anti-Hallucination Mechanisms
Version Evolution and Engineering Discipline
From v1.0 (2026-02) to v3.7.0 (2026-05-05), 15+ releases deepened one theme: making hallucination and structural bias detectable, trackable, and human-interruptible. Highlights include SCR reflection loops (v2.8), style calibration from the user's past papers (v2.9), 29 anti-patterns + 22 IRON RULES (v3.1), PRISMA-trAIce compliance (v3.4), blind sprint contracts (v3.6.2), 18 known hallucination pattern protections (v3.6.7), and plugin packaging with 10 slash commands (v3.7.0).
Engineering practices include lint-driven development (scripts/check_spec_consistency.py, sprint contract tests with 71 test cases, mutation tests for pattern protection) and schema versioning with append-only compliance_history[] for audit integrity.
Notably, the project published a post-publication audit (2026-03-09): all 68 citations verified via web search, finding 21 problematic citations (31% error rate) even after passing three integrity reviews. Corrections followed: 4 fabricated citations removed, 6 author errors and 7 bibliographic details fixed. This honestly demonstrates that integrity gates reduce but do not eliminate errors — and that auditing itself drives improvement.
Feynman-Lens Assessment
Done well: "cargo cult detection" (25 AI高频词 warnings, em-dash limits, three-item-list detection combat AI's patterned writing itself), and teaching agents *how to reason* (Toulmin model, Bradford Hill causal inference) rather than giving steps.
Avoided questions: if base models keep improving, will mandatory human checkpoints become bottlenecks? And the review standards themselves (0-100 scale, ≥80 acceptance) — frame-lock occurs on the human side too. The v3.7.0 README is candid: these optimizations cannot fully solve AI's structural limits — they make limits visible, trackable, and human-interruptible.
Fit
Good for: systematic literature reviews (PRISMA), multi-round revision workflows, strict citation accuracy needs, researchers willing to invest in collaboration, Claude Code users.
Not for: fully automatic paper generation (explicitly rejected), non-Claude-Code workflows (a Codex CLI sibling exists), speed-first use (~$4-6 but heavy time cost per 15k-word paper), commercial use (CC-BY-NC 4.0).