Reference Reading Summarizer

Turn a reference paper into a structured, copyright-safe paper card. This skill answers: What does this paper say?

Do not force every paper into project decisions. Use reference-project-synthesizer after a card exists and the project implication matters.

Skill Directory Layout

<installed-skill-dir>/
├── SKILL.md
├── references/
│   ├── model-routing.md
│   └── reading-modes.md
└── templates/
    ├── paper-card.md
    └── reading-run.md

Core Contract

• Read to the requested purpose, not exhaustively by default.
• Use cheaper models for skim and routine extraction; escalate for central method, theory, benchmark, closest-work, or final-paper claims.
• Preserve provenance: source PDF, pages/sections inspected, reading mode, model tier, confidence, and reviewer status.
• Do not paste long PDF text into cards. Summarize and point to pages/sections.
• Separate paper content from project implications.

Reading Modes

• skim: relevance, role labels, whether deeper reading is needed
• extract-writing: intro framing, paragraph moves, contribution wording, captions, figure/table narration
• extract-method: algorithm, objective, architecture, inference, implementation details
• extract-theory: assumptions, theorem statements, proof ideas, formal definitions
• extract-benchmark: task, dataset, split, metric, protocol, compute, evaluation caveats
• extract-baseline: baseline method role, fairness conditions, comparison requirements
• extract-risk: closest-work threat, novelty boundary, reviewer attack surface
• deep-read: high-value paper where misunderstanding would change project direction

Model Routing

Read references/model-routing.md.

Default:

• Tier 0 deterministic tools: PDF text extraction, metadata, page rendering when available
• Tier 1 cheap sidecar: skim, simple card skeleton, writing-pattern extraction
• Tier 2 normal main model: benchmark/protocol extraction, method extraction, citation support
• Tier 3 strong/deep model: core theory, closest work, must-have baseline, final paper claims

Escalate when the paper is closest work, a core algorithm source, a theory source, a benchmark source, or when the card will support a claim, experiment, baseline, or rebuttal.

Workflow

1. Locate reference/.agent/reference-index.md if available.
2. Identify the paper path and requested reading mode.
3. Create a run artifact under reference/.agent/runs/<run-id>/ when extraction is nontrivial:
   • input-manifest.md
   • extraction-notes.md
   • model.json
   • decision.md
4. Extract only the needed text/pages. Prefer local PDF tools; if unavailable, ask for text snippets or proceed from available metadata and mark limitations.
5. Fill templates/paper-card.md into reference/cards/<paper-id>.md.
6. Update reference/.agent/reading-status.md from unread or skimmed to carded when appropriate.
7. Route:
   • project implications needed -> reference-project-synthesizer
   • broad field map needed -> literature-review-sprint
   • citation coverage needed -> citation-coverage-audit

Paper Card Rules

Every card should include:

• metadata
• reading mode
• role labels
• problem and contribution
• method/theory/benchmark/writing extraction only as relevant
• limitations and uncertainty
• reusable assets
• provenance and confidence

Do not turn cheap skim cards into strong conclusions. Mark confidence: skim or needs-deep-read when appropriate.