Agent Skills
› NeverSight/learn-skills.dev
› reference-reading-summarizer
reference-reading-summarizer
GitHub将项目参考资料（如论文、文档、代码等）转化为结构化的源卡片。通过多种阅读模式提取写作风格、方法、理论、基准及风险等信息，保留来源证据并区分内容与项目影响，为后续决策提供基础素材。
Trigger Scenarios
需要快速浏览和总结论文或PDF报告
整理BibTeX文献或协作者反馈笔记
分析脚本、配置或规范中的实现细节
从多文件包中提取关键信息和关系
Install
npx skills add NeverSight/learn-skills.dev --skill reference-reading-summarizer -g -y
SKILL.md
Frontmatter
{
"name": "reference-reading-summarizer",
"description": "Read and summarize project reference sources into structured source cards. Use for skimming papers, PDFs, Word docs, Markdown notes, BibTeX files, scripts, specs, collaborator feedback, or source bundles; extract writing patterns, methods, theory, benchmarks, baselines, implementation hints, risks, constraints, and project seeds without yet deciding project implications.",
"allowed-tools": "Read, Write, Edit, Bash, Glob, WebSearch, WebFetch",
"argument-hint": "[project-root] [source-id-or-path] [--mode skim|extract-writing|extract-method|extract-theory|extract-benchmark|extract-baseline|extract-risk|extract-feedback|extract-spec|extract-bundle|extract-implementation-hints|extract-project-seed|deep-read]"
}
Reference Reading Summarizer
Turn a reference source into a structured, copyright-safe source card. This skill answers: What does this source say?
Do not force every source into project decisions. Use reference-project-synthesizer after a card exists and the project implication matters.
Skill Directory Layout
<installed-skill-dir>/
├── SKILL.md
├── references/
│ ├── model-routing.md
│ └── reading-modes.md
└── templates/
├── source-card.md
├── paper-card.md
└── reading-run.md
Core Contract
- Read to the requested purpose, not exhaustively by default.
- Use cheaper models for skim and routine extraction; escalate for central method, theory, benchmark, closest-work, final-paper claims, or collaborator decisions.
- Preserve provenance: source path, source type, pages/sections/files inspected, reading mode, model tier, confidence, and reviewer status.
- Do not paste long raw source text into cards. Summarize and point to pages/sections/files.
- Separate source content from project implications.
Source Types
paper-or-pdf: paper PDFs, reports, slide exports, or other PDF referencescollaborator-doc: PDF/Word/text feedback or collaborator-authored notesmarkdown/text-note/latex-source: notes, specs, drafts, or design documentsbibliography: BibTeX collections and citation setsscript/notebook/config-or-spec: implementation hints, protocols, task definitionsbundle: folder containing multiple docs/scripts/bib/configs that should be interpreted as a unit
Reading Modes
skim: relevance, role labels, whether deeper reading is neededextract-writing: intro framing, paragraph moves, contribution wording, captions, figure/table narrationextract-method: algorithm, objective, architecture, inference, implementation detailsextract-theory: assumptions, theorem statements, proof ideas, formal definitionsextract-benchmark: task, dataset, split, metric, protocol, compute, evaluation caveatsextract-baseline: baseline method role, fairness conditions, comparison requirementsextract-risk: closest-work threat, novelty boundary, reviewer attack surfaceextract-feedback: collaborator comments, requested edits, contradictions, decisions, TODOsextract-spec: requirements, constraints, interfaces, acceptance criteria, project assumptionsextract-bundle: bundle inventory, internal relationships, most useful files, missing contextextract-implementation-hints: reusable scripts, configs, APIs, preprocessing, commands, pitfallsextract-project-seed: initial idea, problem, assumptions, resources, open questions, first actionsdeep-read: high-value source where misunderstanding would change project direction
Model Routing
Read references/model-routing.md.
Default:
- Tier 0 deterministic tools: metadata, file inventory, page/text extraction, bundle manifest
- Tier 1 cheap sidecar:
skim, simple card skeleton, writing-pattern extraction, bundle inventory - Tier 2 normal main model: benchmark/protocol extraction, method extraction, collaborator feedback structuring, citation support
- Tier 3 strong/deep model: core theory, closest work, must-have baseline, final paper claims, project seed, collaborator decision conflict
Escalate when the source is closest work, a core algorithm source, a theory source, a benchmark source, a project seed, a collaborator constraint, or when the card will support a claim, experiment, baseline, implementation, paper revision, or rebuttal.
Workflow
- Locate
reference/.agent/source-index.mdfirst; fall back toreference/.agent/reference-index.md. - Identify the source path, source type, and requested reading mode.
- Create a run artifact under
reference/.agent/runs/<run-id>/when extraction is nontrivial:input-manifest.mdextraction-notes.mdmodel.jsondecision.md
- Extract only the needed text/pages/files. Prefer local tools; if unavailable, proceed from available metadata and mark limitations.
- Fill
templates/source-card.mdintoreference/cards/<source-id>.md. For pure papers,templates/paper-card.mdremains a compatible subtype. - Update
reference/.agent/processing-status.mdfromunreadorskimmedtocardedwhen appropriate; updatereading-status.mdif present for compatibility. Memory writeback: this skill writes only toreference/cards/andreference/.agent/; project-level memory writeback (claim/evidence/risk boards) is the job ofreference-project-synthesizerafter synthesis. - Route:
- project implications needed ->
reference-project-synthesizer - broad field map needed ->
literature-review-sprint - citation coverage needed ->
citation-coverage-audit
- project implications needed ->
Source Card Rules
Every card should include:
- metadata and source type
- reading mode
- role labels
- summary or extracted content only as relevant
- limitations and uncertainty
- reusable assets
- provenance and confidence
Do not turn cheap skim cards into strong conclusions. Mark confidence: skim or needs-deep-read when appropriate.
Version History
- e0220ca Current 2026-07-05 21:36
