Paper Analysis

When the PDF is already under paperPDFs

• PDF path pattern: paperPDFs/<Category>/<Venue_Year>/<Year>_<SanitizedTitle>.pdf
• MD path: Replace paperPDFs with paperAnalysis; change extension to .md. Keep the rest unchanged.
• Do not modify an existing .md if the corresponding PDF was not found or could not be read.

When the PDF is not under paperPDFs

1. Infer or ask for: category and venue + year (for example CVPR, 2025 → CVPR_2025).
2. Categories (must choose one): Human_Human_Interaction, Human_Object_Interaction, Human_Scene_Interaction, Motion_Controlled_ImageVideo_Generation, Motion_Editing, Motion_Generation_Text_Speech_Music_Driven, Motion_Stylization.
3. Place PDF: Copy (or move) the PDF to paperPDFs/<Category>/<Venue_Year>/<Year>_<SanitizedTitle>.pdf. Use colocated ingest script if available: python .claude/skills/papers-analyze-pdf/scripts/ingest_pdfs.py <path> --category <cat> --venue <Venue> --year <Year> [--title <Title>]
4. Place MD: Write to paperAnalysis/<Category>/<Venue_Year>/<Year>_<SanitizedTitle>.md.
5. If category or venue/year cannot be inferred, ask user before copying/writing.

Writing principles

Analysis language control

Use analysis_language from the repo-level AGENTS.md as the default for this skill, unless the user explicitly requests a different output language in the current run.

Language rules:

• YAML frontmatter keys always remain in English.
• tags, category, file names, wiki links, and pdf_ref stay English-compatible.
• core_operator and primary_logic values follow the selected analysis language.
• All body prose, bullet explanations, summary text, and evidence text follow the selected analysis language.
• An explicit user request for a specific language overrides the repo default for that run only.

The Three Questions — every analysis must answer these

1. What/Why: What is the true bottleneck? Why solve it now? (data/representation/optimization/controllability/consistency/physical constraints/generalization/evaluation...)
2. How: What key causal knob did the authors introduce? (conditioning / constraint / architecture / objective / sampling / training recipe). What changed → which distribution/constraint/information bottleneck changed → what capability changed?
3. So what: Compared with prior work, where is the capability jump? Which experiment signals best support that claim?

Expression priority

Intuition / structured explanation > mechanism-level abstraction > minimal symbolic notation

• For mechanism, prioritize system-level causality: what changed → which distribution/constraint/information bottleneck changed → what capability changed.
• For evidence, prioritize signal type + conclusion (more controllable / more stable / better long-term consistency / less drift / stronger generalization / less data dependence), rather than piling up numbers/tables.
• Useful high-level framing objects: spaces (input/condition/latent/action/constraint/evaluation space), objects (representation, prior, control variable, decomposition, alignment, retrieval, planning, feedback), processes (what distributions/constraints changed in training; how controllability/stability is injected at inference).

Formula rules

• Do not derive formulas; do not explain loss term-by-term; do not explain every symbol one by one.
• If formulas must be mentioned: keep at most 1-2 verbalized objective/constraint descriptions, attached under the information-flow explanation.
• Only minimal placeholders are allowed (optional): (x) (condition/observation), (z) (latent), (f_\theta) (model), (\mathcal{L}) (objective). They must serve mechanism→capability narration, not derivation.

Content scope

• Keep only the most critical high-level information needed to understand the paper.
• Omit unnecessary detail such as table-by-table restatement, exhaustive hyperparameters, and ablation numbers unrelated to core idea.
• Any needed detail should stay attached to the core thread of "problem–method–capability change" to keep writing short and strong.

Required analysis structure

Each .md must follow this exact layout. If required sections are missing or clearly insufficient, regenerate once. If the paper genuinely cannot fit the structure, keep best attempt, add a one-line note, and tag analysis_mismatch.

1. YAML frontmatter

• Flat only: all keys are top-level scalars or lists; no nested maps.
• For values with colons or long text (for example primary_logic), use quoted strings or | / > multiline scalars.
• Required fields: title, venue, year, tags (list), core_operator, primary_logic, pdf_ref, category, claims (list).
• Frontmatter keys remain English regardless of analysis_language.

---