Paper Compass Learnpath

Paper-Compass-Learnpath

Do one thing: produce an actionable prerequisite learning path before the user reads a paper.

Language Interface

• Supported parameter: lang=zh|en
• Default output language: zh
• If lang is not provided, output in Chinese.
• If lang=en, output all report sections and notes in English.
• If lang=zh, all report section titles must be Chinese.
• Keep technical terms unchanged when translation could reduce precision.

Constraints

C0: Evidence First

• Every prerequisite concept must include at least 1 evidence anchor.
• Evidence format: [Section] "short quote".
• Keep quotes short: at most 25 English words (or similarly short in other languages).
• If direct evidence is missing, mark evidence=indirect and lower confidence.
• Do not put non-evidenced concepts into the Must Learn section.

C1: Prerequisite-Oriented Scope

• Include only concepts required to understand the paper.
• Order by dependency: foundation -> bridge -> paper-specific.
• For each concept, provide star difficulty and a minimum learning goal.
• Star difficulty format:
  • 1 -> ⭐
  • 2 -> ⭐⭐
  • 3 -> ⭐⭐⭐
  • 4 -> ⭐⭐⭐⭐
  • 5 -> ⭐⭐⭐⭐⭐

C2: Personalization First

• If user provides memory=<path>, read that file first.
• If not provided, try ~/Documents/know/memory.md.
• If the file does not exist, continue and mark memory as not loaded.
• Do not reteach mastered knowledge unless needed for paper-specific deltas.

C3: Honesty First

• Explicitly write 信息不足 (for zh) or insufficient information (for en) when needed.
• Never fabricate sections, quotes, or resource links.
• Use low-confidence when certainty is limited.

C4: Impact and Venue Verification

• For ALL papers: Must fetch venue and impact information via API.
• NEVER claim "preprint" or "unpublished" without verification.
• For recent arXiv papers (within past 2 years): These are HIGH PRIORITY for venue verification:
  • Many arXiv papers get published at conferences (NeurIPS, ICLR, ICML, ACL, etc.) within months
  • Conference cycles: NeurIPS (Dec), ICLR (May), ICML (Jul), ACL (May-Aug)
• Impact data: Fetch via Semantic Scholar API (Bash + curl):
  • Citation count
  • Venue name
  • TLDR (one-line summary)
• If API fails: Mark as venue待验证 or citations待验证, NOT "preprint/unpublished"

C5: Output Structure Compliance

• Section 0 (论文快照) must follow this exact structure:

  - 标题: {title}
  - 作者: {authors}
  - 年份: {year}
  - 发表信息与venue: {venue_name} | JCR 分区: {Q1/Q2/Q3/Q4/N/A} | CCF 等级: {A/B/C/N/A}
  - 来源: {paper_url_or_path}
  - **影响力**: {citation_count_and_awards_if_known_or_search_online}
  

• Section 6 (30 分钟快速起步) must only contain:
  • 关键实验结论: {1-3 sentences summarizing key findings}
  • No numbered steps, no formulas, no extra content
• Section 7 must be ## 7. **Sources**: followed by reference links

Input Normalization

arXiv Paper Access Strategy (Multi-Source Fallback)

arXiv resources (PDF, HTML, abs) are PRIMARY - always use first:

1. arXiv API (abs): Fetch https://export.arxiv.org/api/query?id_list={id} → get title, authors, abstract, year
2. Stable PDF download with retry: use curl with redirect-following, retry, timeout, and a browser-like user agent
3. HTML fallback: if PDF still fails, fetch https://arxiv.org/html/{id} when available
4. Semantic Scholar API (supplement): curl -s "https://api.semanticscholar.org/graph/v1/paper/ARXIV:{id}?fields=venue,citationCount,tldr" → get venue, citations

Priority Order:

Priority 1: arXiv API → metadata (title, authors, abstract, year)
Priority 2: Stable PDF download + Read tool → full paper content (sections, quotes, methods)
Priority 3: arXiv HTML fallback → recover readable sections when PDF fails
Priority 4: Semantic Scholar API via curl → venue, citations, TLDR (supplement)

CRITICAL:

• NEVER use WebSearch - it only works in US and will fail in other regions
• For arXiv full-text access, prefer direct arXiv/API access first; semantic-scholar is only a metadata supplement, not the primary full-text path
• NEVER use WebFetch for arxiv.org - domain verification will block it; use curl/download instead
• arXiv PDF is the BEST source for full paper content, but downloads may intermittently fail; always use retry before giving up
• Prefer https://arxiv.org/pdf/{id} over hardcoding .pdf suffix when downloading
• A successful PDF download must leave a non-empty local file before continuing
• NEVER download to /tmp or other system temp paths; always save files under the current working directory in ./papers/
• When using Read, always read ./papers/{id}.pdf or ./papers/{id}.html from the current workspace so Windows path resolution does not fail

Workflow

Step 1: Fetch Paper Metadata and Content

Use Bash + Python/curl for arXiv access and full-text retrieval (no WebSearch, no WebFetch for arxiv.org):

# 1. Get arXiv metadata
python3 -c "
import urllib.request, xml.etree.ElementTree as ET, json
NS = 'http://www.w3.org/2005/Atom'
url = 'https://export.arxiv.org/api/query?id_list=ARXIV_ID'
with urllib.request.urlopen(url, timeout=30) as r:
    root = ET.fromstring(r.read())
# Extract title, authors, abstract, year, pdf_url...
"

# 2. Always download to the current workspace under ./papers/
mkdir -p papers
curl -L --retry 5 --retry-delay 2 --retry-all-errors \
  --connect-timeout 15 --max-time 120 \
  -A "Mozilla/5.0" \
  -o ./papers/ARXIV_ID.pdf https://arxiv.org/pdf/ARXIV_ID

# 3. Verify local PDF exists and is non-trivial
test -s ./papers/ARXIV_ID.pdf

# 4. If PDF failed, try HTML fallback
curl -L --retry 3 --retry-delay 2 --retry-all-errors \
  --connect-timeout 15 --max-time 60 \
  -A "Mozilla/5.0" \
  -o ./papers/ARXIV_ID.html https://arxiv.org/html/ARXIV_ID

# 5. Get Semantic Scholar metadata (venue, citations)
curl -s "https://api.semanticscholar.org/graph/v1/paper/ARXIV:ARXIV_ID?fields=venue,citationCount,publicationVenue,tldr"

Then read content:

# Prefer Read tool on ./papers/ARXIV_ID.pdf
# If the PDF download failed but HTML exists, parse ./papers/ARXIV_ID.html instead
# Do not read /tmp/... paths on Windows

Extract and record:

• Title, authors, year (from arXiv API)
• Publication metadata:
  • venue name (from Semantic Scholar API via curl)
  • JCR quartile (journal-only; otherwise N/A)
  • CCF rank (if applicable; otherwise N/A)
• Impact data (from Semantic Scholar API via curl):
  • Citation count
  • TLDR (one-line summary)
• Section titles and quotes (prefer PDF via Read tool, fallback to HTML)
• Key areas: method, experimental setup, critical appendix details

When download or APIs fail:

• If arXiv API fails, stop claiming verified title/year metadata
• If PDF fails but HTML works, continue with lowered confidence for quote extraction
• Mark section-level evidence as 信息不足 if neither PDF nor HTML is accessible
• Mark venue/citations as 待验证 if Semantic Scholar API fails
• Still produce a valid report with available information
• If direct Semantic Scholar API lookup is weak or ambiguous, you may use /semantic-scholar only as a metadata cross-check for venue, citations, DOI, and TLDR

Step 2: Load User Prior Knowledge (memory.md)

Read references/memory-format.md, then classify user knowledge into:

• mastered
• familiar
• basic
• unknown

If a concept is not present, default to unknown.

Step 3: Extract Prerequisite Candidates

Use these signals:

• Core modules, operators, and training strategies in method sections
• Required prior models/theory (e.g., Transformer, contrastive learning, quantization)
• Experimental settings that materially affect conclusions

Tag each candidate:

• role=must: essential to understand the main method
• role=bridge: helps with key details
• role=optional: useful but not required

Step 4: Bind Evidence and Section Usage

For each must/bridge concept, bind at least one evidence item:

• section: section name or number
• quote: short original quote
• usage: one-sentence explanation of usage in this paper

If the concept appears across multiple sections, include multiple evidence anchors.

Step 5: Build Learning Order and Difficulty

Topologically sort by dependencies to produce order.

Difficulty levels:

• 1: term-level, ~30-60 minutes
• 2: standard module-level, ~1-2 hours
• 3: mechanism reasoning-level, ~2-4 hours
• 4: implementation/theory detail-level, ~4-8 hours
• 5: cross-paper synthesis-level, >8 hours

For each concept provide:

• minimum_goal: what "good enough" means
• estimated_time: suggested time investment

Step 6: Prune with Memory

• mastered and not central to this paper's novel delta: downgrade to skip/review-optional
• familiar: keep a minimal refresher path
• basic/unknown: keep in primary path

Explicitly state:

• Which concepts are skipped and why
• Which familiar concepts are still kept due to paper-specific novelty

Step 7: Recommend Resources (Paper + Video)

For each must concept recommend:

• At least 1 paper/survey link
• At least 1 video link (lecture/talk/tutorial)

Follow references/resource-sourcing.md.

When lang=zh:

• Keep original paper links as mandatory references.
• Add more Chinese-learning links when relevant:
  • bilibili lectures/tutorials
  • Zhihu columns/answers with technical depth
  • CSDN posts only when they are implementation-useful and not low-quality copy.
• Anthropic official docs/blog articles can be included when they clarify core concepts.
• Prefer high-signal Chinese resources over generic summaries.

Step 8: Generate Report

Select template by language:

• lang=zh -> references/template.zh.md
• lang=en -> references/template.en.md
• fallback -> references/template.md

Read selected template and write:

• File name: {timestamp}--paper-compass-learnpath-{short-title}__learnpath.md
• Path: current working directory (./)

After writing, report the absolute output path to the user.

Output Quality Checklist

• Every Must Learn item includes section-grounded evidence.
• Learning order is executable (no dependency inversion).
• Difficulty and time estimates are internally consistent.
• Resource links are valid and concept-relevant.
• Personalization clearly explains memory-driven differences.

Source: cenzihan/paper-compass-skill — distributed by TomeVault.