Awesome LLM Paper Wiki

name: paper-wiki description: | Academic literature management and survey system. Supports journal organization, tag management, literature survey report generation, and submission recommendation. Trigger: user mentions "literature", "paper", "survey", "journal", "submission", "paper-wiki", or requests scan/report/recommendation on an initialized vault.

paper-wiki — Academic Literature Survey Skill

Manage a local Markdown literature vault. Scan, organize, tag, survey, and recommend — all from one skill.

What This Skill Does

paper-wiki turns a folder of paper Markdown files into a structured, indexed literature vault with:

• Journal organization: auto-sort papers into paper/{direction}/{journal_abbr}/
• Tag management: multi-dimensional tagging (task, method, dataset, domain, signal, etc.)
• Survey reports: journal reports, direction reports, method/dataset stats, idea novelty surveys, literature reviews
• Submission guidance: journal scoring, revision suggestions, resubmit audits, review loops

Quick Start

• "Initialize vault" → init
• "Scan Battery papers" → scan-organize
• "Generate RESS journal report" → journal-report
• "Recommend submission target for my paper" → submission-recommend

Configuration

Configuration is read from config.json at project root. Key sections: output_lang, templates, research_workflows, web_search. For the full schema, field descriptions, and defaults, see docs/configuration.md. Before workflows that depend on vault paths, language settings, or report behavior, inspect config.json and apply its active values.

Output Language Rules

All internal content (this SKILL.md, schemas, scripts, templates) is in English. User-facing output follows output_lang from config.json: zh → Chinese, en → English. File paths, directory names, variable names, and frontmatter keys remain in English regardless of output_lang.

Report Citation Policy

All report-generation workflows must maintain a citation registry:

• Mark every evidence-bearing paper with [N] in prose, tables, and claims. Number by first appearance; reuse numbers.
• The final References section must include every cited paper and ONLY cited papers.
• Reference format: [N] Title. Journal, Year. DOI: xxx. URL: xxx. Source: local path or [arxiv-web].
• Omit missing fields (DOI, URL, year) rather than using placeholders.
• Before finalizing: verify citation numbers are continuous, every in-text citation has one reference entry, and every reference entry is cited at least once.
• Direction-review scope: standard 40-80 refs; deep 80-120 refs. Every reference must be cited in the body.

Full-Coverage Report Policy

For journal-report and direction-report, the evidence pipeline is mandatory. The protocol enforces screening, coverage ledger, citation traceability, and unsupported-claim risk checks. report_family.py --complete is the authoritative completion gate.

Full protocol: Load references/evidence-validation-protocol.md together with the report workflow reference file before execution.

Summary: Stage 1 builds screening/ledger evidence files → Stage 2 writes the report with full-coverage matrix → Stage 3 verifies count equality → Stage 4 runs the hard completion gate. --metadata-only overrides to the legacy deterministic path.

How Workflows Execute

SKILL.md is the dispatch layer, not the operational manual. When a user request matches a row in the routing table below:

1. Load the referenced workflow file(s) from the skill directory.
2. Execute the steps defined in that file.
3. Do not execute a workflow from the routing table alone — catalog entries provide routing, not procedure.
4. If a workflow references another workflow by name, load that workflow's reference file before using it.

Required files may include both a workflow reference file and supporting reference files. Load all listed files before execution.

Workflow Routing

Route user intent to the appropriate workflow and reference file:

Routing notes:

• Every workflow runs independently. If preconditions are unmet, output a clear message but do NOT auto-chain unless the user explicitly requests "full pipeline".
• direction-review = literature review / survey synthesis. idea-survey = novelty/similarity analysis for a specific idea.
• For reviewer-backed workflows (idea-evidence, idea-create, idea-discover, idea-claim-novelty-check): prefer Codex-compatible MCP reviewer, otherwise dual-agent, otherwise degraded single-agent. Label degraded mode.
• For review/audit workflows (auto-review-loop, resubmit-audit, paper-review-loop): prefer Codex-compatible MCP reviewer. In Claude Code, explicitly call codex. Fallback to dual-agent then single-agent degraded.
• Conference Seed Mode: When idea-create is triggered with an explicit conference paper file path, load references/workflows/19-idea-evidence.md and execute its ### Conference Seed Mode section first. The routing table entry for idea-create includes this path; the activation gate is in references/workflows/20-idea-create.md.
• MCP-first web search: For web-find and web-digest workflows, MCP tools (paper-search-mcp, arxiv-mcp-server) are preferred when available. Fallback to CLI (web_search.py) when MCP unavailable. At workflow completion, if CLI was used, output MCP installation suggestion. Do NOT auto-install MCP servers; follow official docs if user agrees.

Precondition Matrix

Workflow Catalog

Template System

Templates live in templates/generic/. Domain-specific templates in templates/domains/{Direction}/. The template registry is tracked in config.json → templates.registry.

Status check (status workflow): reports template coverage and staleness. Lint (lint workflow): detects missing or stale domain templates. Regeneration threshold is set in config.json → templates.regeneration_threshold.

Safety Rules

1. Never delete files in paper/ — these are the user's original paper Markdown files
2. File moves require --dry-run first, then user confirmation before --apply
3. Preserve ## User Notes — never overwrite content under this heading in any file
4. Log everything — all file operations go to workspace/logs/
5. Tag priority — user tags > rule tags > manually confirmed extra suggestions (never remove user tags)
6. No cross-direction moves — files stay within their research direction unless user explicitly requests
7. No overwrites — if target file exists during organize, log as conflict, do not overwrite

Frontmatter Template

---
title: ""
authors: []
journal: ""
journal_abbr: ""
published_date: null
doi: ""
url: ""
source_path: ""
direction: ""
canonical_type: "research" | "survey" | "review"

tags_task: []
tags_method: []
tags_dataset: []
tags_domain: []
tags_signal: []
tags_application: []
tags_metric: []
tags_custom: []

status: "unread"
reading_priority: "medium"
updated_at: timestamp
---

# Title

## Source

## Abstract

## Keywords

## User Notes
<!-- User-maintained section. Scripts must never overwrite this. -->

Source: moonlarry/awesome-llm-paper-wiki — distributed by TomeVault.