reference-corpus-analyzer
GitHub将多篇文献转化为结构化对比矩阵，支持方法差异、性能差距及趋势分析。适用于撰写相关工作、识别核心竞品及发现研究空白，需基于已有的源卡片进行，区别于单篇总结或项目关联技能。
Trigger Scenarios
Install
npx skills add NeverSight/learn-skills.dev --skill reference-corpus-analyzer -g -y
SKILL.md
Frontmatter
{
"name": "reference-corpus-analyzer",
"description": "Produce a multi-paper comparison matrix across a literature corpus with tiered read depth. Use when multiple papers need to be compared side-by-side for method differences, performance gaps, closest-work ranking, or trend identification — distinct from per-paper source cards (reference-reading-summarizer) and single-paper project linking (reference-project-synthesizer).",
"allowed-tools": "Read, Write, Edit, Bash, Glob, WebSearch, WebFetch",
"argument-hint": "[project-dir] [--corpus <path>] [--top-n <N>] [--mode compare|rank|trend|gap]"
}
Reference Corpus Analyzer
Synthesize a literature corpus into a structured comparison matrix. This skill answers: across these N papers, who does what, how do they differ, and where is the open space?
Use this skill when:
- related-work writing requires a side-by-side method comparison across multiple papers
- you want to identify the closest 3–5 papers and understand exactly how they differ from each other
- a literature survey needs trend identification across publication years or venues
- you want to map open gaps across a set of existing approaches before writing the related-work section
- you have 5+ source cards and want a comparison table rather than individual summaries
Do not use this skill to create per-paper source cards — use reference-reading-summarizer for that. Do not use this skill to link a paper to your project's claims — use reference-project-synthesizer for that. Use this skill after source cards exist; avoid re-reading raw PDFs unless a card is insufficient.
Pair this skill with:
reference-reading-summarizerupstream: produce source cards before running corpus analysisreference-project-synthesizerupstream or downstream: link individual papers to project memory before or after comparisonrelated-work-positioning-writerdownstream: use the comparison matrix to write novelty-boundary paragraphsbaseline-selection-auditdownstream: use the ranking to identify must-have baselinesliterature-review-sprintwhen the corpus is not yet assembled and a broader topic survey is needed first
Skill Directory Layout
<installed-skill-dir>/
├── SKILL.md
└── templates/
└── comparison-matrix.md
Progressive Loading
- Read
reference/cards/to find available source cards before reading raw sources. - Read
reference/.agent/source-index.mdorreference/.agent/reference-index.mdto get the corpus inventory. - Read
memory/claim-board.mdwhen the comparison should be anchored to specific project claims. - Read
templates/comparison-matrix.mdbefore writing the output matrix.
Core Principles
- Tiered depth: read deeply only the top-N closest papers; skim the rest for placement.
- The comparison matrix is a project artifact, not a free-text essay — it should be queryable.
- Ranking closest work requires a clear criterion: task overlap, method overlap, or claim overlap.
- Gaps should be stated specifically: "no paper does X under constraint Y" beats "X is underexplored".
- Do not invent paper properties not stated in the source card or the paper itself.
Step 1 — Assemble the Corpus
Read reference/.agent/source-index.md (or reference-index.md) to list available sources.
For each source, record:
- source ID and title
- card availability:
has-card,no-card,partial-card - initial relevance estimate:
core,related,background,tangential
If cards are missing for sources that appear highly relevant, route to reference-reading-summarizer first.
Step 2 — Select Tiered Read Depth
Assign read depth to each source:
| Tier | Sources | Read depth |
|---|---|---|
| Deep | Top 3–5 closest by task + method overlap | Full source card; re-read raw source if card is insufficient |
| Standard | Next 5–10 related works | Source card only |
| Survey | Remaining background papers | Title + abstract + card summary |
Criterion for "closest": same task, same claim type, overlapping method family, or shared benchmark.
Step 3 — Build the Comparison Matrix
Read templates/comparison-matrix.md.
Dimensions to compare (select those relevant to the project):
- Task / problem: what problem does the paper address?
- Method family: what is the core mechanism (attention, diffusion, RL, prompting, etc.)?
- Key innovation: what is the single thing this paper claims to contribute?
- Benchmark / dataset: what is it evaluated on?
- Primary metric: what metric is reported?
- Best reported result: the headline number (with venue and year for context)
- Limitations acknowledged: what does the paper say it cannot do?
- Relationship to our work: closer / complementary / orthogonal / superseded by ours
For each tier-1 (deep) paper, also add:
- Closest claim to ours: the specific claim that most overlaps with our paper's contribution
- Key differentiator: in one sentence, how our work differs from this paper
Save to reference/corpus-analysis-<date>.md.
Step 4 — Rank Closest Work
Produce a ranked list of the top-5 closest papers with:
Rank: 1
Paper: [title] ([venue year])
Overlap: task=high / method=medium / claim=high
Closest claim: [their specific claim that overlaps ours]
Differentiator: [one sentence: how we differ]
Novelty risk: high / medium / low
Reviewer action: cite as closest work / cite as baseline / cite as background
A paper with novelty risk: high and method=high overlap is the paper whose related-work paragraph needs the clearest boundary statement.
Step 5 — Identify Gaps and Trends
Gaps: what combinations of (task, method, constraint, benchmark) are not yet addressed by any paper in the corpus?
Gap: no paper addresses [X] under [constraint Y] with [method family Z]
Evidence: papers A, B, C address X but not under Y; papers D, E address Y but not X
Opportunity: our work fills this gap by [brief description]
Trends (optional, for survey mode):
- Which method families are gaining / losing papers over the last 3 years?
- Which benchmarks are becoming standard vs. falling out of use?
- What claims were controversial 2 years ago and are now accepted?
Step 6 — Write Memory Writeback
- Update
reference/.agent/source-index.mdwith read-depth assignments - Update
memory/risk-board.mdfor any high-novelty-risk closest-work findings - Update
memory/claim-board.mdif the comparison changes the novelty framing of a claim - The comparison matrix itself is saved in
reference/— do not copy it intomemory/
Final Sanity Check
Before finishing:
- tier-1 papers have been read at full card depth
- the comparison matrix has consistent dimensions across all papers
- top-5 closest-work ranking has differentiators in one sentence each
- gaps are stated specifically (not as vague "future work" language)
- novelty-risk papers are flagged for
related-work-positioning-writer
Version History
- e0220ca Current 2026-07-05 21:36
