A Claude Skill that turns academic papers into structured research understanding and implementable innovation points — with every claim tagged by evidence strength, so the analysis never quietly upgrades a guess into a finding.
Status: beta · Version: v0.5.5-beta · License: MIT
It reads a paper the way a careful reviewer does: separating what the paper actually claims from what you might infer, marking formulas, tables, and figures it could not parse as unavailable rather than guessing, and refusing to fabricate a bibliographic field it cannot verify.
- Structured paper decomposition — identity, problem, method, datasets, metrics, baselines, results, and evidence coverage.
- Claim/inference separation — paper claims stay distinct from the model's own reasoning.
- Real novelty assessment — evaluates what is actually new instead of restating the abstract.
- Limitation mining — converts weaknesses into concrete research opportunities.
- Implementable innovation points — each with experiments, baselines, metrics, risks, and a feasibility score.
- Multi-paper synthesis — cross-paper research directions, positioning, and innovation frameworks.
- Evidence-aware asset reading — formulas, framework diagrams, charts, tables, references, multi-column layout, OCR text, and supplementary material, each with an uncertainty label.
- Safety gates — Evidence Strength, Claim Safety (including Chinese trigger words), Table Identity, Ablation Table Integrity, Bibliographic Verification, SOTA Claim Classification, and a three-state Quality Audit (pass / partial / fail).
- GB/T 7714-2015 citations by default, with incomplete fields marked missing rather than invented.
- Optional Codex Adapter for repository-level maintenance, report review, and PR workflows.
paper input (PDF / text / extracted assets)
-> identity + problem + method parsing
-> evidence coverage assessment [unavailable assets marked, not guessed]
-> claim / inference separation
-> novelty + limitation analysis
-> innovation point generation [experiments, baselines, metrics, risks, feasibility]
-> safety gates [evidence · claim · table · bibliographic · SOTA]
-> quality audit [pass / partial / fail]
-> report
Behavior depends on how many papers you supply:
| Input | Output |
|---|---|
| One paper | Paper-centered deep analysis; innovation points are single-paper extensions |
| Multiple papers | Literature synthesis and innovation framework by default — not N separate full reports |
For multiple papers the Skill first builds an inventory and evidence cards, then uses only papers with sufficient evidence. Weakly related or poorly parsed papers are marked as background or excluded. Separate full reports are produced only on explicit request.
# Personal skill — available across all projects
mkdir -p ~/.claude/skills
cp -r paper-innovation-analyst ~/.claude/skills/
# Project skill — available only in the current repository
mkdir -p .claude/skills
cp -r paper-innovation-analyst .claude/skills/Then invoke it in a prompt:
Use paper-innovation-analyst to analyze this paper and propose implementable innovation points.
Build a clean release archive, then upload the .skill file through Claude's Skills settings:
python scripts/package_skill.py .
# -> dist/paper-innovation-analyst-v0.5.5-beta.skill
# -> dist/paper-innovation-analyst-v0.5.5-beta.zipWorks with Claude products that support custom Skills, including Claude Code and Claude.ai.
The Skill reads papers supplied through Claude's own document handling, uploaded PDFs, pasted text, or other document-reading Skills. When you want local extraction instead:
pip install -r requirements-optional.txt # OCR also needs a system Tesseract install
python scripts/extract_paper_assets.py path/to/paper.pdf --out outputs/paper_assets --ocr-mode autopython -m py_compile scripts/validate_skill.py scripts/extract_paper_assets.py scripts/package_skill.py
python scripts/validate_skill.py .
python -m pytest tests -q
python scripts/package_skill.py .paper-innovation-analyst/
├── SKILL.md # Primary instruction source: the 8-phase workflow
├── CHANGELOG.md
├── CLAUDE.md
├── requirements-optional.txt # Optional deps for local PDF extraction
├── references/ # Rules the Skill consults while analyzing
│ ├── review-rubric.md
│ ├── domain-addenda.md
│ ├── document-ingestion-pipeline.md
│ ├── cv-detection-addendum.md
│ ├── gbt7714-2015-examples.md
│ └── idea-quality-gates.md
├── templates/ # Output shapes the Skill fills in
│ ├── paper-analysis-template.md
│ ├── innovation-brief-template.md
│ ├── experiment-plan-template.md
│ ├── multi-paper-comparison-template.md
│ └── reviewer-report-template.md
├── examples/example-prompts.md
├── scripts/ # Validation, extraction, packaging
├── tests/
├── codex/ # Optional Codex Adapter (see below)
└── .github/workflows/validate.yml
references/ and templates/ carry the two halves of the Skill's behavior: references/ holds the rules and gates applied while reading a paper, templates/ holds the report structures those rules produce. SKILL.md is the entry point that ties them together.
The repository ships two entry points. The Claude Skill (SKILL.md) is the primary instruction source; the Codex Adapter (codex/) is a lightweight repository-level layer for the OpenAI Codex CLI. The adapter is NOT a standalone .skill installation package and is not a replacement for the Claude Skill — it references the full Skill rather than duplicating it.
| Claude Skill | Codex Adapter | |
|---|---|---|
| Entry point | SKILL.md |
AGENTS.md and codex/AGENTS.md |
| Best for | Deep single-paper analysis, synthesis, innovation mining, experiment design | Repository maintenance, report review, template improvement, testing, PR workflows |
| Output | Full detail, complete formula/table/figure expansion | Reports too, but must not over-compress |
# Option A: copy the adapter into your own project alongside SKILL.md
cp -r codex/ /path/to/your/project/
# Option B: use this repository directly
codexCodex CLI reads codex/AGENTS.md as project instructions, which references SKILL.md for the full workflow. codex/prompts/ holds task-specific rules; codex/checklists/ holds the verification gates. Both platforms share the same core safety rules listed under Features.
- Not certified across every Claude product or PDF type. The repository has local validation and pytest coverage; it has not been formally certified beyond that.
- Capability boundary — complex documents may parse partially. Document asset extraction is best-effort: it does not guarantee stable parsing of every complex PDF, scanned page, mathematical formula image, figure, chart, table, multi-column reading order, or supplementary package. Anything unreadable is marked unavailable or uncertain — never guessed.
- Batch mode is for synthesis, not depth. Batch outputs are research planning drafts and need human verification before use in proposals, theses, or publications.
MIT — see LICENSE.
Version history: CHANGELOG.md.