summary - SKILL.md Agent Skill

name: summary description: This skill should be used when providing concise, priority-first summaries of research papers. Use for quickly understanding what a paper does, what is genuinely new, how it works, what the main results are, and whether the limitations materially weaken the conclusion.

Paper Summary Skill

Generate priority-first summaries of research papers with keyword extraction. Read ingested paper markdown, write a concise and decision-useful summary to notes.md, and backfill extracted keywords into YAML frontmatter tags.

When to Use This Skill

Summarizing a research paper after ingestion for quick understanding
Extracting the paper's main claim, novelty, method, and main results
Preparing compact research notes for later recall
Generating keywords for indexing and retrieval
As a follow-up step after paper-ingestion

Pipeline Context

Typical workflow: paper-ingestion -> summary -> paper-translate / paper-validator

Input: A paper directory produced by paper-ingestion, containing full_text.md (and optionally full_text_zh.md after translation).

Default Behavior

By default, produce a priority-first full-paper summary only.

The summary should help the reader answer, as quickly as possible:

What does this paper do?
What is the main claim?
What is genuinely new?
What is the motivation behind the method design?
How does the method actually work?
What are the main results?
Do the limitations materially weaken the conclusion?

Do not produce section-by-section summaries.

Language Settings

Default output language: Chinese (中文)
Can be specified as English by the user
Chinese mode rules:
- Technical terms and proper nouns remain in English when appropriate
- Paper titles remain in original language
Keywords are always in English regardless of output language

Reading Strategy

Do not summarize the paper in section order.

Use a priority-first reading strategy:

Read abstract, introduction, conclusion, and the main result tables/figures first
Identify the core claim, novelty, motivation, method shape, and strongest support
Drill into method, main experiments, and ablations only as needed
Allocate summary space by importance, not by section order

By default, do not spend summary budget on:

References
Appendix
Prompt details
Long tables item-by-item
Related work survey details
Future work, unless it clarifies an important limitation

Workflow

Step 1: Locate Paper Files

User provides a paper directory path or full_text.md path
Confirm the directory contains full_text.md
Check if full_text_zh.md exists (for keyword backfill only)

Step 2: Read Paper Content

Read full_text.md only
Do not use full_text_zh.md as the source for summarization

Step 3: Generate Summary

Write a concise priority-first summary using the format below.

The summary should emphasize importance, novelty, motivation, mechanism, and main support, not uniform coverage.

Step 4: Extract Keywords

Extract at least 4 keywords from the paper (typically 4-8)
Follow the keyword rules below
Include keywords in the summary output

Step 5: Write Summary to notes.md

Save the complete summary output into notes.md in the same paper directory
Replace placeholder note content if present

Step 6: Backfill Keywords to Tags

Append extracted keywords to the tags: list in the YAML frontmatter of full_text.md
If full_text_zh.md exists, append the same keywords there too
Do not modify any content outside YAML frontmatter

Output Format

# Paper Summary: {Title}

## Takeaway
[1-3 sentences. State what the paper does and why it matters.]

## Core Claim
[1-3 sentences. State the main claim or thesis of the paper.]

## What's New vs Prior Work
[1-4 sentences. Explain the genuine novelty and what is stronger or different than recent work.]

## Method
[Usually 2-4 short natural paragraphs, not a bullet list. Explain the method from first principles.
Start from the problem pressure, failure mode, or design motivation that makes the method necessary.
Then explain the core idea, the main components or stages, the key objective or optimization target if important,
the training or inference flow when central, and why these design choices should produce the claimed gains.
This should usually be the most detailed section of the summary.
Prefer causal explanation and paper storyline over module inventory.
It is acceptable to use multiple paragraphs if that makes the logic clearer.
Be detailed but not verbose, and omit appendix-only variants or low-importance implementation detail.]

## Key Results
[2-5 sentences. Summarize the most important results and the main comparison points.
Focus on the outcomes that matter most for the paper's claim; do not list every benchmark.]

## Limitations
[1-2 sentences. Keep this brief, and note whether the limitations materially weaken the main conclusion.]

## Historical Impact (conditional — include ONLY when criteria in Summary Requirements are met)
[2-5 sentences. Describe how this work influenced subsequent research:
what paradigm it established, what methods or techniques it introduced that became widely adopted,
or what research directions it opened. Focus on concrete downstream impact, not speculation.]

## Keywords
keyword1, keyword2, keyword3, keyword4, ...

Summary Requirements

Priority-first: Allocate space by importance, not by section order
Decision-useful: Help the reader quickly understand the claim, novelty, motivation, mechanism, and whether the paper is convincing
Method-first detail: Method should usually be the most detailed section in the summary
First-principles method: Explain the method from motivation -> core idea -> mechanism -> why it should work
Narrative clarity: Preserve the paper's storyline; explain what problem pressure leads to each major design choice
Mechanism over inventory: Explain how the method works and why the main design choices matter; do not turn Method into a catalog of components
Multiple paragraphs allowed: Use 2-4 short paragraphs in Method when needed for clarity; prefer natural prose over bullet inventories
Concise: Do not rewrite the full paper in compressed form
Technically accurate: Preserve precise terminology and comparisons
Novelty-focused: Clearly separate what the paper does from what is actually new
Results-focused: Summarize the strongest outcomes, not every experiment
Historical Impact gate: Include the Historical Impact section ONLY when ALL of: (a) you have confident knowledge that this paper's method, framework, or idea was adopted or extended by multiple subsequent works, (b) you can name concrete examples of the downstream influence (named methods, paradigm shifts, standard components). Omit the section entirely when: the paper is very recent (published within ~6 months), is an incremental improvement on existing methods, or you lack sufficient knowledge about its downstream influence. When in doubt, omit.

Keyword Extraction Guidelines

Quantity: At least 4, typically 4-8
Coverage: Prefer a mix of domain, method, technique, and task keywords
Priority: Paper's own keywords > abstract concepts > core terms from full text
Format: English, lowercase except proper nouns and standard abbreviations
No spaces in tags: Use hyphens (-) for multi-word tags

Examples:

test-time-training
diffusion-transformer
single-cell-analysis
kernel-engineering

Keyword Backfill Rules

Backfill means appending extracted keywords into the tags: list in the YAML frontmatter of full_text.md (and full_text_zh.md if it exists). Insert keywords before the aliases: line.

If tags: already contains entries beyond paper, assume keywords were already backfilled and skip.

Do not modify content outside YAML frontmatter.

Important Constraints

Do not introduce information not present in the source text
Do not editorialize beyond evidence-based judgment about whether limitations weaken the conclusion
Do not summarize the paper section-by-section
Do not spend space on appendix/reference material unless essential to the core claim
Maintain technical precision appropriate for an academic audience
Only backfill keywords when producing a full paper summary