Paper Reading Notes Generator

Workflow

Step 1: Identify Mode and Read the PDF / Existing Note

There are two valid entry modes:

• Draft mode: create a new reading note from a PDF
• Audit mode: review an existing note, especially a note already labeled Pass 2

In Audit mode, read the existing note first and treat the Pass 2 label as untrusted until the audit confirms it.

First, read the PDF skill for file handling:

view /mnt/skills/public/pdf/SKILL.md

Then locate the original PDF. Prefer the repo copy under papers/source/ when available. If the user uploaded a PDF for this turn, use that copy.

Extract the full text from the PDF. Use pdfplumber for text extraction as it preserves layout better:

import pdfplumber

with pdfplumber.open("/mnt/user-data/uploads/<filename>.pdf") as pdf:
    full_text = ""
    for i, page in enumerate(pdf.pages):
        text = page.extract_text()
        if text:
            full_text += f"\n--- Page {i+1} ---\n{text}"

Also extract tables separately — they contain key experimental results:

    all_tables = []
    for i, page in enumerate(pdf.pages):
        tables = page.extract_tables()
        for j, table in enumerate(tables):
            all_tables.append({"page": i+1, "table_index": j, "data": table})

If you are re-reading the PDF for quote verification or Pass 2 audit, keep a clean extracted text file under the same topic's papers/ tree so future quote checks do not require fresh extraction.

Preferred location:

<topic>/papers/text/<pdf-stem>_clean.txt

Rules for the clean text file:

• Keep only the readable text needed for accurate quotation and page lookup
• Preserve page markers such as --- Page N ---
• Remove obvious extraction noise when safe to do so (duplicated blank lines, broken line wraps, repeated headers/footers if clearly identifiable)
• Do NOT save multiple noisy variants like _layout, _raw, _debug unless the user explicitly asks
• If papers/text/ does not exist, create it

Step 2: Read the Note Template, Audit Checklist, and Style Guide

view /mnt/skills/paper-reading-notes/references/note-template.md
view /mnt/skills/paper-reading-notes/references/quality-checklist.md
view /mnt/skills/paper-reading-notes/references/style-guide.md

The template defines the output structure. The checklist defines whether the note can remain Pass 1, be promoted to Pass 2, or be downgraded from an existing Pass 2 label after audit. The style guide defines the note's positioning and language contract. Read all three before writing.

Step 3: Fill the Template

Work through the template section by section. Here are the extraction strategies for each:

Meta Information

• Title, Authors, Venue: Extract from the first page header/footer area. If venue isn't clear from the PDF, note "Venue not identified in PDF — check manually."
• arXiv ID: Look for patterns like arXiv:XXXX.XXXXX in headers or footers.
• Citation count: Write "Check Semantic Scholar" — don't guess.
• Read date: Use today's date.
• Priority & Reading progress: Default to P1 and Pass 1. The user will adjust.

One-line Summary

Format: [System/Method name] [does what] on [which benchmarks], achieving [key metric improvement over baseline]. Extract from Abstract + Conclusion. Be specific with numbers.

Problem Setting

This is critical — find the exact quotes where the authors state:

1. The problem they solve (usually Introduction paragraphs 1-3)
2. Why existing approaches are insufficient (usually Introduction paragraphs 3-5 or Related Work)
3. Their assumptions (often implicit — check the experimental setup)

Always include Section references and page numbers: (Section 1, p.2)

Core Method

• Write the overview in your own words, explaining the architecture/pipeline
• For the Design Choices table: identify every major architectural decision. Check if each one has a corresponding ablation experiment. This is crucial — unablated design choices are potential gaps.

Key Results

• Extract from the main results table(s). Identify the strongest baseline for each benchmark.
• Calculate Δ (delta) if not provided: this_method - strongest_baseline
• For ablation findings: identify which component removal causes the largest drop

Limitations

• Author-stated: Search for "limitation" section (usually near the end) and "future work"
• Observed: Look for:
  • Benchmarks that seem too easy or too narrow
  • Baselines that are missing (recent strong systems not compared)
  • Assumptions that might not hold in practice
  • Scalability concerns not addressed

Gap Signals

Search the paper for:

• Phrases like "we leave X to future work", "beyond the scope", "does not address"
• Experimental results where performance drops significantly in specific conditions
• Comparisons where certain categories of methods are absent

Each gap signal MUST have a direct quote or data point with location.

In Audit mode, do not rewrite everything blindly. First identify which claims are decision-critical for Pass 2:

• core problem / insufficiency quotes
• key design choices and ablation status
• author-stated limitations
• observed limitations
• gap signals
• connections to other papers / KB update readiness

Only reopen and patch the sections whose evidence chain is weak, missing, or suspicious.

Step 4: Positioning and Language Convention

Treat the reading note as a research intermediate artifact, not as:

• final survey prose
• a generic summary
• a temporary TODO draft

Follow this language contract:

• Template structure / headers / table headers: English
• Direct quotes / paper terminology / benchmark names / model names / metric names: English
• Analytical commentary / observations / research linkage: Chinese

Default rule:

Chinese for analysis, English for evidence fidelity.

Required style constraints:

• Do not write mixed Chinese-English slang sentences
• Do not switch terminology for the same concept within one note
• Do not write Relation to My Research like a generic survey paragraph
• Do not keep unresolved placeholders in notes that are meant to support KB sync

If uncertainty remains:

• write 原文未明确说明
• or write 当前判断基于 Table X + Section Y 的间接证据

instead of leaving vague placeholders or guesses.

Example of correct style:

## Limitations
- **Author-stated limitations**: "Our method requires pre-defined action spaces and cannot handle novel UI elements" (Section 6, p.9)
- **My observed limitations**:
  1. 该方法只在 Android 平台验证，跨平台泛化仍未证明。
  2. 实验任务平均步骤数较短，因此对长序列任务的支持证据不足。
- **Experimental design gaps**: 缺少与更强 recent baseline 的统一重跑对比。

Step 5: Run the Pass 2 Audit

After the draft note is filled, run the checklist in:

view /mnt/skills/paper-reading-notes/references/quality-checklist.md

Apply these rules strictly:

• New notes start as Pass 1 unless the audit proves otherwise.
• Existing notes labeled Pass 2 must also pass the same audit again; do not trust the old label by default.
• Do NOT mark a note as Pass 2 just because every section in the template is non-empty.
• Upgrade to Pass 2 only if every Hard Gate passes and the note is strong enough to support comparison-matrix / gap-tracker updates without reopening the PDF.
• If any Hard Gate fails, keep or change the note to Pass 1 and record the blocking items for the user.
• If a claim depends on inference rather than direct evidence, keep the note at Pass 1 unless the inference is clearly labeled and the supporting quote/data is still present.

When auditing an existing Pass 2 note, use a skeptical standard:

• If a direct quote cannot be verified quickly from the note alone, re-check the PDF
• If page numbers or section references look uncertain, re-check the PDF
• If a gap signal is strong enough to influence comparison-matrix / gap-tracker, re-check the PDF unless the note already contains exact supporting quote/data
• If the note cannot support a survey paragraph without reopening the PDF, it is not truly Pass 2

Typical blockers for Pass 2:

• missing section/page references in core evidence fields
• gap signals that are mostly inference, not quote/data-backed
• limitations that only restate the authors' own wording
• research linkage too vague to place into the survey or KB
• direct quotes that cannot be confidently traced back to the original PDF

Necessary conclusion for a true Pass 2 note:

• the paper can be written into a survey paragraph directly from the note
• comparison-matrix and gap-tracker can be updated from the note without reopening the PDF
• core claims are fully supportable from the paper's original wording or result tables

Step 6: Output

Save the completed note to: /mnt/user-data/outputs/<short-name>-notes.md

Where <short-name> is a lowercase-hyphenated version of the paper's short name or system name (e.g., appagent-notes.md, awm-notes.md).

After saving, use present_files to share with the user.

Step 7: Post-output Reminder

After presenting the file, remind the user:

笔记已生成。

- 如果这是 `Pass 1` 草稿：请重点检查 `My observed limitations`、`Relation to My Research` 和 `Gap Signals` 是否需要进一步收口。
- 如果这是 `Pass 2` 笔记：它应已可直接支持 KB 同步，不应再残留未收口占位。

Also report the audit result in plain language:

• If promoted: Pass 2 audit: passed + 2-3 reasons it passed
• If not promoted: Pass 2 audit: not yet + the blocking checklist items
• If re-auditing an old Pass 2 note that fails: Pass 2 audit: failed on re-check + whether the note was downgraded to Pass 1
• If the PDF was reopened: report that the original was re-checked and include the saved clean text path

Quality Criteria

Minimum quality checks before delivery:

• One-line summary contains: subject + method + benchmark + number
• Problem setting has direct quotes with Section + page references
• Design Choices table is filled; ablation verification column is marked
• Limitations include at least one observation NOT stated by the authors
• Gap signals have original text evidence (Section + page), not just inference
• Key Passages section has 1-3 critical quotes with exact locations
• Language follows the rule: Chinese analysis + English evidence anchors
• Research-facing sections are aligned with the current main line / gap tracker rather than generic commentary

For Pass 2 decisions, use the full audit in references/quality-checklist.md rather than only this short list.

Important Notes

• Do NOT hallucinate page numbers or section references. If you cannot locate the exact position, write "位置待确认" instead of guessing.
• Do NOT copy the abstract as the method overview. Rewrite in your own words.
• Do NOT leave the Design Choices table empty — even a 2-row table is better than none. If the paper doesn't clearly describe design choices, note what you can infer and mark them as "从实验设置推断".
• If table extraction from the PDF is poor (garbled text), note this and suggest the user manually fill in the Key Results table.
• The priority of this skill is evidence extraction accuracy. A shorter note with correct page references beats a longer note with vague attributions.
• Pass 2 is a readiness judgment, not a formatting judgment. A complete-looking note can still be only Pass 1.
• For Pass 2 re-audits, prefer false-negative over false-positive. If evidence feels shaky, re-open the PDF and verify.
• In finalized notes, prefer fully written analytical judgments over ⚠️ NEEDS YOUR INPUT. Placeholders are acceptable only for genuine unresolved researcher decisions, not as the default output style.