Agent Skills
› NeverSight/learn-skills.dev
› reference-library-manager
reference-library-manager
GitHub管理项目参考库，负责扫描、索引、去重和监控论文及文档的处理状态。不用于深度阅读或分析，而是维护基础设施层面的源文件元数据与进度跟踪。
Trigger Scenarios
扫描参考目录结构
生成或更新源文件索引
检查处理状态和重复项
执行批量元数据清理
Install
npx skills add NeverSight/learn-skills.dev --skill reference-library-manager -g -y
SKILL.md
Frontmatter
{
"name": "reference-library-manager",
"description": "Manage project reference sources under reference\/. Use when scanning, ingesting, indexing, deduplicating, monitoring, or tracking processing status for papers, PDFs, Word docs, Markdown notes, BibTeX files, scripts, specs, or source bundles without deeply reading them.",
"allowed-tools": "Read, Write, Edit, Bash, Glob",
"argument-hint": "[project-root] [--scan] [--ingest] [--monitor] [--status]"
}
Reference Library Manager
Manage the project reference/source library as durable research infrastructure. This skill answers: What sources do we have, and what is their processing state?
Do not use this skill for deep source understanding or project implications. Use reference-reading-summarizer for "what does this source say?" and reference-project-synthesizer for "what does this source mean for our project?"
Skill Directory Layout
<installed-skill-dir>/
├── SKILL.md
├── scripts/
│ └── scan_reference_library.py
├── references/
│ └── library-policy.md
└── templates/
├── reference-index.md
└── reading-status.md
Core Contract
- Default project reference root is
<project-root>/reference/. - A source can be a paper PDF, collaborator PDF/Word doc, Markdown note, BibTeX file, script, spec/config, notebook, webpage snapshot, or hand-built folder bundle.
- Raw source files may be private, licensed, large, or collaborator-provided. Do not assume they should be committed to a public repo.
- Commit derived indexes, source cards, and project-use notes only when they are sanitized and useful.
- Prefer deterministic scanning for filenames, hashes, size, and modification time.
- Use cheap sidecars only for fuzzy title matching or metadata cleanup, not for raw source storage.
Expected Project Layout
reference/
├── sources/ # non-paper sources and bundles, often ignored/private
│ ├── collaborator-docs/
│ ├── markdown/
│ ├── bundles/
│ └── misc/
├── papers/ # backward-compatible paper/PDF location
├── cards/ # stable source cards from reference-reading-summarizer
├── project-use/ # project implications from reference-project-synthesizer
├── summaries/ # optional summaries
├── notes/ # human notes
└── .agent/
├── source-index.md
├── reference-index.md # compatibility alias
├── processing-status.md
├── reading-status.md # compatibility alias
├── metadata-gaps.md
├── duplicate-check.md
└── runs/ # raw scan/read trajectories, usually ignored
Model Routing
- Tier 0 script: scan files/folders, hashes, size, mtime, missing cards, processing-status skeletons.
- Tier 1 cheap sidecar: fuzzy duplicate titles, filename-to-title cleanup, coarse source type and role tags.
- Tier 2/3 models: not needed here; route to the reading or project synthesis skills.
Workflow
- Locate project root and
reference/. Preferreference/sources/plus backward-compatiblereference/papers/; if neither exists, scan supported source files recursively underreference/. - Read
references/library-policy.md. - Run the scanner when indexing is needed:
python3 <installed-skill-dir>/scripts/scan_reference_library.py --project-root .
- Review generated or updated files:
reference/.agent/source-index.mdreference/.agent/reference-index.mdreference/.agent/processing-status.mdreference/.agent/reading-status.mdreference/.agent/metadata-gaps.mdreference/.agent/duplicate-check.md
- Mark each source's intended role when known:
idea-seedwriting-exemplarmethod-sourcetheory-sourcebenchmark-sourceimplementation-sourcecollaborator-feedbackproject-specbaselinecitation-supportclosest-workreviewer-risk
- Route next action:
- unread but important ->
reference-reading-summarizer - card exists and project decisions are needed ->
reference-project-synthesizer - missing citation coverage in a draft ->
citation-coverage-audit - broad field search ->
literature-review-sprint
- unread but important ->
Output Rules
- Store raw scanner trajectories under
reference/.agent/runs/only when useful. - Store durable index state under
reference/.agent/. - Store stable summaries as source cards under
reference/cards/. - Never paste long raw source text into memory or public docs.
- Report only counts, major gaps, duplicates, and recommended next sources unless the user asks for detail.
- Memory writeback: update
reference/.agent/processing-status.mdwhen sources are added, deduplicated, or change status. Project-level memory writeback (claim/evidence/risk boards) is handled byreference-project-synthesizerafter synthesis, not by this skill.
Version History
- e0220ca Current 2026-07-05 21:36
