Add paper-writing skill

.agents/skills/paper-writing/SKILL.md

@@ -0,0 +1,71 @@
+---
+name: paper-writing
+description: Write CS/AI papers with progressive disclosure. Use when drafting or revising sections such as abstract, introduction, related work, method, figures, experiments, or rebuttal text. Starts from a top-level workflow and loads only the section-specific reference needed. Includes arXiv source-fetch workflow for mining LaTeX organization from exemplar papers.
+---
+
+# Paper Writing
+
+## Mission
+
+Provide a section-aware paper-writing workflow that is:
+
+1. evidence-first
+2. progressively disclosed
+3. LaTeX- and structure-friendly
+4. optimized for CS and AI papers
+
+Do not load all writing guidance at once. Start here, then open only the reference file needed for the current section.
+
+## Default Workflow
+
+1. identify paper type and draft stage
+2. load `references/workflow.md`
+3. load only the section file needed right now
+4. if related-work style mining is needed, load `references/arxiv-source-workflow.md`
+5. use `scripts/fetch_arxiv_source.py` to inspect exemplar paper sources
+6. write only what is supported by actual evidence, experiments, and citations
+
+## Loading Map
+
+Open these references on demand:
+
+1. `references/source-index.md`
+   - external sources and what each one informs
+2. `references/workflow.md`
+   - overall writing order, collaboration rules, and section contracts
+3. `references/abstract.md`
+   - abstract structure and compression rules
+4. `references/introduction.md`
+   - problem framing, gap, contributions, and section logic
+5. `references/related-work.md`
+   - literature grouping, synthesis, comparison writing, and citation density
+6. `references/arxiv-source-workflow.md`
+   - how to download and inspect arXiv LaTeX sources
+7. `references/method.md`
+   - method narrative, notation, pipeline order, and algorithm exposition
+8. `references/architecture-figures.md`
+   - architecture drawing heuristics and code-based tooling
+9. `references/experiments.md`
+   - evaluation logic, ablations, robustness, and error analysis
+10. `references/tables-and-layout.md`
+   - result tables, layout, appendix splitting, and figure-table coordination
+
+## Writing Rules
+
+1. Write from claims backward:
+   - claim
+   - evidence
+   - section role
+   - prose
+2. Prefer synthesis over surface summary.
+3. One paragraph should usually do one job.
+4. Keep novelty claims proportional to the real delta.
+5. For figures and tables, prefer code-generated assets that can be regenerated.
+6. For related work, inspect representative arXiv source when organization matters.
+
+## Scripts
+
+1. `scripts/fetch_arxiv_source.py`
+   - download and unpack an arXiv source bundle
+   - list candidate main `.tex` files and section fragments
+   - useful for `Related Work`, `Method`, and layout mining

.agents/skills/paper-writing/references/abstract.md

@@ -0,0 +1,56 @@
+# Abstract
+
+## Job
+
+Compress:
+
+1. problem
+2. gap
+3. method
+4. evidence
+5. implication
+
+The abstract is not a motivational essay.
+
+## Default Structure
+
+Use 4-5 sentences:
+
+1. exact problem context
+2. concrete bottleneck in prior work
+3. method and key mechanism
+4. strongest result with numbers where possible
+5. why the result matters
+
+## Sentence Rules
+
+Sentence 1:
+
+- name the problem directly
+- avoid empty openers like "Recent advances..."
+
+Sentence 2:
+
+- specify the actual failure mode or gap
+- avoid generic "however" claims
+
+Sentence 3:
+
+- name the method and 1-2 essential mechanisms
+- do not enumerate every module
+
+Sentence 4:
+
+- report concrete results
+- say what the comparison is against
+
+Sentence 5:
+
+- state the takeaway, not sales copy
+
+## Rewrite Checklist
+
+1. remove hype words
+2. replace adjectives with mechanism or metric
+3. cut background better placed in introduction
+4. keep tense consistent

.agents/skills/paper-writing/references/architecture-figures.md

@@ -0,0 +1,75 @@
+# Architecture Figures
+
+## Job
+
+Architecture figures should clarify the method faster than prose can.
+
+## Design Rules
+
+1. show information flow, not decoration
+2. keep the reader's path obvious
+3. use 3-6 major blocks, not every implementation detail
+4. align names with the method section exactly
+5. annotate the paper's novel block or changed path
+
+## Recommended Figure Workflow
+
+1. sketch the semantic blocks first
+2. decide whether the figure is:
+   - pipeline
+   - dataflow
+   - training loop
+   - system deployment
+3. draft quickly in code
+4. only then polish spacing, color, and labels
+
+## Code-Based Tooling Choices
+
+### TikZ
+
+Use when:
+
+1. the paper is LaTeX-native
+2. you need publication-grade vector consistency
+3. the diagram is stable enough to justify fine control
+
+Strengths:
+
+1. integrates directly into LaTeX
+2. precise typography and alignment
+3. easy to keep consistent with paper fonts
+
+### Graphviz
+
+Use when:
+
+1. you need fast iteration on graph structure
+2. you want automatic layout before manual polishing
+3. the figure is mostly nodes and edges
+
+Strengths:
+
+1. fast text-based editing
+2. automatic layout
+3. good for dependency and pipeline drafts
+
+### Practical Recommendation
+
+Default pattern:
+
+1. draft with Graphviz or a simple text diagram if structure is unclear
+2. finalize in TikZ when the figure is stable and will appear in the paper
+
+## Labeling Rules
+
+1. use noun phrases for blocks
+2. use verbs on arrows only when action matters
+3. avoid paragraphs inside boxes
+4. make repeated flows visually repetitive on purpose
+
+## Figure Review Checklist
+
+1. can a reviewer retell the method from the figure
+2. is the novelty visually identifiable
+3. can the figure survive grayscale or print
+4. does the figure introduce names not used in the text

.agents/skills/paper-writing/references/arxiv-source-workflow.md

@@ -0,0 +1,51 @@
+# arXiv Source Workflow
+
+## Why Use Source Instead Of PDF Alone
+
+Source bundles expose:
+
+1. section splits
+2. table files
+3. figure code or assets
+4. appendix organization
+5. macros and naming conventions
+
+This is especially useful for:
+
+1. related work organization
+2. method subsection design
+3. experiment/table layout reuse
+
+## Default Workflow
+
+1. pick exemplar papers by section quality, not just citation count
+2. download source with `scripts/fetch_arxiv_source.py`
+3. inspect candidate main `.tex` and included subfiles
+4. read only the relevant source fragments
+5. extract organization patterns, not sentences
+
+## Example
+
+`ReAct` (`2210.03629`) was verified during skill creation:
+
+1. arXiv e-print works
+2. the source bundle contains split files such as:
+   - `text/method.tex`
+   - `table/*.tex`
+   - figure assets
+
+That makes it a good reference for section decomposition and table/figure separation.
+
+## Command
+
+```bash
+python3 .agents/skills/paper-writing/scripts/fetch_arxiv_source.py 2210.03629
+```
+
+## What To Look For
+
+1. does the paper isolate related work or blend it into introduction
+2. how many method subsections it uses
+3. where tables live
+4. whether architecture figures are code-generated or image-based
+5. how appendix references are wired

.agents/skills/paper-writing/references/experiments.md

@@ -0,0 +1,59 @@
+# Experiments
+
+## Job
+
+Experiments should answer whether the paper's claims hold.
+
+## Default Order
+
+1. setup and implementation details
+2. main results
+3. ablations
+4. sensitivity or robustness
+5. qualitative or error analysis
+
+## Core Principle
+
+Every experiment must map to a claim.
+
+Before writing a subsection, identify:
+
+1. claim being tested
+2. baseline needed
+3. metric that decides success or failure
+
+## Main Results Section
+
+1. start with the primary benchmark or setting
+2. compare against the strongest relevant baselines
+3. explain the pattern before diving into every number
+
+## Ablations
+
+1. remove or vary one component at a time
+2. explain what the ablation isolates
+3. do not present ablations as a random checklist
+
+## Robustness / Sensitivity
+
+Use when the paper claims:
+
+1. generalization
+2. stability
+3. low-resource effectiveness
+4. scale behavior
+
+## Error Analysis
+
+Use error analysis when:
+
+1. the method claims interpretability
+2. failure modes are central to the paper
+3. aggregate metrics hide behavior differences
+
+## Writing Rules
+
+1. explain the evaluation logic, not just the metric names
+2. separate implementation detail from result interpretation
+3. highlight negative or null results when informative
+4. say when a baseline is omitted and why

.agents/skills/paper-writing/references/introduction.md

@@ -0,0 +1,53 @@
+# Introduction
+
+## Job
+
+The introduction should make a reviewer answer:
+
+1. what exact problem is being solved
+2. why current approaches are insufficient
+3. what this paper changes
+4. why the evidence should be believed
+
+## Default Paragraph Order
+
+1. problem and stakes
+2. concrete gap or bottleneck
+3. paper idea and intuition
+4. evidence preview
+5. contribution bullets
+6. optional section roadmap
+
+## Paragraph Rules
+
+Paragraph 1:
+
+- define the task and stakes precisely
+- avoid field-history narration unless it matters
+
+Paragraph 2:
+
+- show where prior approaches break
+- use 1-2 crisp examples or failure patterns
+
+Paragraph 3:
+
+- present the paper's idea at the right abstraction level
+- do not dump full method detail
+
+Paragraph 4:
+
+- preview strongest evidence
+- name datasets, environments, or metrics when possible
+
+Contribution bullets:
+
+- each bullet should be independently defensible
+- separate method, evaluation, and systems contributions
+
+## Common Failure Modes
+
+1. spending too much space on broad motivation
+2. hiding the novelty until late
+3. contribution bullets that are just tasks completed
+4. promising more than experiments support