wiki-pdf-ingest

name

wiki-pdf-ingest

description

**ONLY** triggered when the user explicitly wants to **import a paper PDF into their personal wiki knowledge base** (not for generic PDF reading / summarizing / translating). Extracts the PDF into a deep-read markdown note under `raw/科研/论文资料/` and can chain into wiki-ingest to produce a wiki/papers/ brief. [STRICT triggers]: 'add this PDF to wiki / knowledge base', 'import pdf to wiki', 'wiki-pdf-ingest', 'ingest this pdf into wiki', 'add paper pdf to knowledge base', 'deep-read this PDF into my wiki'. [DO NOT trigger]: generic 'read this PDF' / 'summarize PDF' / 'translate PDF' / 'convert PDF to PPT' / paper-review use cases — those belong to anthropic-skills:pdf / paper-mkdocs-site / paper-reviewer respectively.

wiki-pdf-ingest — PDF Paper → Personal Wiki

Trigger threshold (strict)

Only trigger in these situations:

1. User explicitly requests a wiki-scoped PDF import (e.g. "add this PDF to wiki", "import paper pdf to my knowledge base", "ingest pdf into wiki")
2. User drops a PDF AND the conversation is clearly continuing wiki-related work (previously discussed wiki/ directory, currently doing ingest, organizing papers/, etc.)

Do NOT trigger for (use other tools/skills instead):

• Generic PDF reading → Read tool / anthropic-skills:pdf
• PDF summary / translation → regular conversation
• PDF → slide / PPT → paper-figure-to-pptx / paper-mkdocs-site
• PDF merge / split / edit → anthropic-skills:pdf
• Peer-review critique on the PDF → paper-reviewer

Disambiguation default behavior

If user's request is ambiguous between "read PDF" and "import into wiki" (e.g. "help me look at this PDF"):

• Do NOT auto-trigger this skill
• Ask back: "For this PDF do you want: (a) quick read of content (I'll just read) / (b) deep-read and add to wiki knowledge base (trigger wiki-pdf-ingest) / (c) peer-review critique / (d) other?"

Core Function

Automatically convert a paper PDF into a deep-read markdown note in the user's wiki style, saved to raw/科研/论文资料/YYYY-MM-DD-<author><year>-<short-slug>.md. Optionally trigger wiki-ingest to produce a wiki/papers/ brief.

Reference format

Deep-read notes follow the style of existing high-quality examples in raw/科研/论文资料/. Read one for calibration, e.g. any 1000+ line file.

Mandatory skeleton

Every deep-read note MUST include these H2 sections, in order:

# <Paper Title>

## 文章信息 / Article Metadata
- **Title**: <full>
- **Authors**: <author1, author2, ... or "X et al.">
- **Venue**: <journal/conference + year>
- **DOI**: <10.xxxx/...>
- **Keywords**: <from paper>

## 摘要整理 / Abstract Digest

### Research Background
<1-2 paragraphs, paraphrased>

### Research Goal
<1 paragraph, what the paper claims to solve>

### Main Findings
<1-2 paragraphs, core contributions>

## 论文框架结构 / Paper Structure

<chapter-by-chapter, 2-5 sentences each>

## 核心技术 / Core Techniques

### System Model
<variable definitions + objective function + constraints>

### Core Formulas
<LaTeX preserved + physical meaning; full derivation chain, not just conclusions>

### Algorithm
<pseudocode or step-by-step>

### Key Design Tradeoffs
<which parameters/structure are knobs, their roles>

## 创新点 / Innovation Points

<3-5 contributions claimed by authors, extracted verbatim>

## 实验 / Experiments

### Simulation Setup
<hardware, baseline methods, scenarios, metrics>

### Key Results
<core comparison data>

### Ablation Studies
<if any>

## 个人反思 / Personal Reflection (AI draft, user to refine)

<!-- TODO (user): refine personal angle -->
- Why read this (logical linkage to my work): <TBD>
- Method transfer points: <AI draft based on related wiki concepts>
- Assumptions/limitations: <AI draft based on paper limitations>
- Baseline alignment: <TBD>

## References
- DOI: <10.xxxx/...>
- <other URLs / Zotero link / arXiv ID if present>

Workflow

Step 0: Pre-check PDF

1. Confirm PDF exists via the user-provided path
2. Check size: if > 50 pages, use Read tool's pages parameter to process in chunks (max 20 pages each)
3. Detect type: paper / survey / tech report / slide deck — slide decks aren't suitable input, warn user

Step 1: Extract PDF metadata + content

Prefer anthropic-skills:pdf skill if loaded; otherwise use the Read tool (supports PDF up to 20 pages).

Chunked reading strategy for large PDFs:

• Chunk 1: pages 1-5 → title, authors, abstract, keywords, intro opening
• Chunk 2: pages 6-10 (or Intro + Related Work) → related work
• Chunk 3: pages 11-15 (or System Model + Method) → core techniques
• Chunk 4: pages 16-20 (or Experiment) → experiments
• Final chunk: pages N-5 ~ N → conclusion + bibtex (if last page)

Step 2: Structured extraction

Fill per-section per the skeleton above:

1. Article Metadata: extract from first-page metadata. DOI is priority. venue typically appears in the copyright line ("0000-0000 © IEEE" followed by venue name for IEEE papers).
2. Abstract Digest: translate/paraphrase the paper's abstract into 3 paragraphs (background / goal / findings). Don't copy verbatim — paraphrase while preserving key claims.
3. Paper Structure: scan all H1/H2 titles + 2-5 sentence summary per section.
4. Core Techniques: for the heavy chapters:
   • System model → variables + objective + constraints (complete, no pruning)
   • Formulas → preserve LaTeX + add commentary
   • Algorithm → convert to step-by-step or pseudocode
5. Innovation Points: prefer extracting the author's own contribution list (usually in Intro or a Contributions subsection).
6. Experiments: baselines, metrics, key numbers — replicate essential tables.

Step 3: Personal reflection auto-draft (constrained)

⚠️ Academic framing: follow the "Academic-framing criteria for ideas" in wiki-ingest SKILL.md.

Draft 4 reflection items (user refines later), similar to wiki-ingest's paper-scaffold style:

1. Logical linkage (not "personal motivation"): how the paper connects logically to the user's existing wiki concepts / syntheses / territories.
2. Method transfer points: which component could transfer to which concept/work the user has.
3. Assumptions/limitations: does the paper's implicit assumption hold in the user's typical scenarios?
4. Baseline alignment (if user has a matching task): can the paper's baselines serve as reference?

Each item gets a <!-- TODO (user): ... --> marker for review.

Step 4: Name + write + verify

1. File naming: YYYY-MM-DD-<author><year>-<short-slug>.md
   • YYYY-MM-DD = today's date (ingest date, not paper publication date)
   • <author> = first author surname (lowercase)
   • <year> = paper publication year
   • <short-slug> = 3-5 English words, kebab-case
   • Example: 2026-04-20-wang2025-satellite-aav-priority.md
2. Write to: raw/科研/论文资料/<filename>.md (don't touch the original PDF)
3. Verify:
   • wc -l (expected 300-1500)
   • grep to confirm all mandatory sections present
   • Show first 30 lines for user quick-review

Step 5: (Optional) Trigger wiki-ingest

Ask the user:

"Generated raw/科研/论文资料/<filename>.md (N lines). Continue with wiki-ingest to create wiki/papers/<paper_id>.md brief?"

If yes:

• Call wiki-ingest skill (paper branch)
• Apply cp + Edit strategy to generate paper brief
• Update wiki/index.md + log.md
• Trigger post-ingest idea-suggestion hook

Token budget

• Small PDF (< 20 pages): ~15-25K tokens
• Medium PDF (20-50 pages, surveys / long papers): ~30-50K tokens
• Large PDF (> 50 pages): recommend user pre-split

Fallback: scanned PDFs / image-heavy PDFs

• If PDF is scanned (image-only), Read cannot extract text
• Prompt user: use OCR tool first, or manually paste abstract

Common failure modes

1. DOI not extractable: some arXiv preprints lack DOI; leave blank or use arXiv ID
2. Formula extraction fails: complex figures may produce garbled LaTeX; ask user to hand-verify
3. Table extraction fails: complex-layout tables may lose structure; mark <!-- TODO: manually supplement Table X -->
4. Author name parsing error: Chinese PDFs or very long author lists may overflow; fall back to "X et al."

Collaboration with wiki-ingest

This skill handles PDF → raw/*.md (steps 1-4). wiki-ingest handles raw/*.md → wiki/papers/ + concept/entity extraction (the rest).

They couple loosely through the raw/科研/论文资料/ directory. User can:

• Only call this skill (generate deep-read notes, don't enter wiki)
• Chain this skill + wiki-ingest (one-shot from PDF to wiki/papers/)

Git convention

Files generated by this skill:

• raw/ new files go to git stage (raw is the immutable audit trail)
• Do NOT commit the original PDF itself (usually too large; gitignore or raw/assets/ but not in git)
• Commit message: pdf-ingest: <short paper title>