Multi-Perspective Academic Paper Review Simulator
Simulates a full journal peer-review process with 5 independent reviewers and an editorial synthesis into a decision and revision roadmap.
Evidence shows Iron Rule #7 explicitly treats submitted manuscripts, review comments, etc. as untrusted data, prohibiting embedded instructions from altering agent identity, routing, tool use, network/API calls, file writes, disclosure rules, or workflow constraints, demonstrating awareness against prompt injection; Iron Rule #6 imposes a read-only constraint on not modifying the manuscript, but there is no least-privilege, user confirmation (only after Phase 0 shows configuration card), data-flow transparency, rollback mechanism, or sensitive-data handling details. Deductions for lack of external-effect control, unclear permission scope, and partial confirmation.
SKILL.md describes detailed orchestration workflow and multiple modes, but references numerous reference files and scripts (e.g., scripts/check_pipeline_integrity.py) not present in the provided files, relies on external prompt sources (reviewer_sprint_prompt_source.md) and cross-skill routing rules, and key paths cannot be confirmed from static files. Failure feedback mentions exist (e.g., [PROTOCOL-VIOLATION] tags), but no actual execution or test evidence. Deductions for inability to reproduce and missing dependencies.
SKILL.md provides rich trigger keywords (in Chinese, English, Korean), non-trigger scenario tables, mode selection guides, clear scenarios and expected outputs, fitting academic paper review needs. However, capability boundaries are not explicit (e.g., unsupported paper types or fields), and environment fit: core functionality relies on Claude Code and external models, no mention of mainland China network adaptation. Deductions for incomplete boundary descriptions and insufficient environment fit evidence.
SKILL.md has clear version (1.10.0), last-updated date, status, and metadata, well-structured, provides quick start and integration relationships. But lacks install/dependency notes, known-limitation disclosure (only partially in DA agent), troubleshooting guide, complete changelog, and clear maintenance responsibility. License metadata is NOASSERTION, provenance unclear. Deductions for incomplete versioning governance and unclear license.
The skill's goal is clear: simulate 5 reviewers and provide editorial decision and revision roadmap, with template support for output format, claiming to streamline peer review. But no actual output samples or third-party execution evidence, marginal value claims (e.g., time savings) unverifiable, core task completion questionable. Deductions for lack of execution evidence and verifiable outputs.
SKILL.md cites references (e.g., Kim et al. 2026) and internal scripts, but provides no external verification or test results. Calibration mode claims to measure FNR/FPR, but no implementation details or output samples. No third-party evidence in static files, key claims cannot be independently verified. Deductions for lack of external corroboration and reproducible experimental evidence.
- The skill relies on Claude Code and external online resources; core functionality may not work in mainland China network environments.
- License metadata is NOASSERTION, provenance unclear; verify copyright and redistribution terms before use.
- The skill claims to have test scripts and CI, but this review is only based on static files, without verifying their actual existence and effective operation.
- The review process involves multi-agent and cross-model calls, which may incur costs; users should understand the cost structure.
- The skill's output depends on model judgment, which may be inaccurate or biased; manual review is recommended.
What does this skill do, and when should you use it?
This skill simulates the complete international journal peer review process: it automatically identifies the paper's field, dynamically configures five reviewers (Journal-Fit Reviewer, three peer reviewers, and a Devil's Advocate) who review from five non-overlapping perspectives, and then uses a separate editorial synthesizer to produce a structured Editorial Decision and Revision Roadmap. It supports full review, re-review (verification), quick assessment, methodology focus, Socratic guided, and calibration modes. It emphasizes evidence-driven review, prohibits fabricating comments, and enforces a strict read-only constraint so reviewers never modify the manuscript.
Reads the full paper text, uses a field analyst agent to identify the discipline and methodology type, and dynamically generates identities for five reviewers. The five reviewers independently produce review reports from different angles: journal fit, methodology rigor, domain contribution, cross-disciplinary perspective, and core-argument challenges (Devil's Advocate). The editorial synthesizer consolidates the five reports, identifies consensus and disagreements, arbitrates disputes, flags Devil's Advocate critical issues, and produces an editorial decision letter and a prioritized revision roadmap. In re-review mode, it uses a traceability matrix to verify whether revisions addressed first-round comments. Calibration mode measures the reviewer's false negative/positive rates against a gold set.
- Before submission, an author wants a comprehensive evaluation of their paper's quality and a prediction of what editors might decide, using full mode.
- After receiving revision requests, an author needs to check whether their revisions adequately address reviewer concerns, using re-review mode.
- When an author only wants a quick overall quality assessment and highlights of key issues, they can use quick mode.
- When a paper has serious methodological or statistical problems, an author wants an in-depth review of research design and statistical analysis, using methodology-focus mode.
- When an author wants to discover issues in their paper themselves through guided Socratic dialogue, they can use guided mode.
- When an author doubts the reliability of the reviewer's scoring and wants to measure its error profile before trusting the scores, they can use calibration mode.
What are this skill's strengths and limitations?
- Deep simulation of real journal review with five non-overlapping perspectives, including a dedicated Devil's Advocate to challenge core arguments.
- Produces an official editorial decision letter and a prioritized revision roadmap via the synthesizer, directly usable for subsequent revisions.
- Emphasizes evidence-based review: every finding is tied to a specific location, and vague comments are prohibited.
- Offers re-review mode with a traceability matrix to verify that revisions are actually addressed, avoiding rubber-stamp outcomes.
- Includes calibration mode to measure the reviewer's own error rates, enhancing trustworthiness.
- Not licensed for commercial use (CC-BY-NC 4.0).
- No public test suite or benchmark to objectively measure review quality.
- Complex implementation with many internal agents and protocols; may be overkill for simple quick assessments.
- Some advanced features (cross-model verification, model tiering) rely on Claude Code environment and may not transfer to other platforms.
How do you install this skill?
This skill is part of the academic-research-skills repository. In Claude Code, it can be installed via the plugin marketplace: /plugin marketplace add Imbad0202/academic-research-skills then /plugin install academic-research-skills. Alternatively, clone the repository manually and place the academic-paper-reviewer folder into your project's .claude/skills or global skills directory.
How do you use this skill?
After installation, simply say to Claude Code: "Review this paper: [paste paper or provide file]" to trigger a full review. The output includes five reviewer reports and an editorial decision. For other modes, you can say "quick review", "check methodology", "guide me", "verification review", or "calibrate reviewer". The skill supports English, Traditional Chinese, and Korean trigger phrases.