paper-digest

name	paper-digest
description	Create or revise structured PaperDigest records from scientific papers, PDFs, PDF URLs, extracted paper text, abstracts, or paper metadata. Use when Codex needs to turn one paper into a reusable digest asset for literature review, paper ingest, wiki filing, citation-backed synthesis, or batch paper processing.

Paper Digest

Produce a reusable structured understanding of one paper. A digest is an intermediate compiled asset, not a wiki page.

This skill is schema-bound. Before changing digest behavior, read references/digest-schema.md and preserve the named contract there. When changing caching, persistence, PDF capability handling, or failure recovery, also read references/runtime-behavior.md.

Examples

• User: "digest this PDF for literature review" -> use paper-digest.
• User: "create a structured digest for this arXiv paper" -> use paper-digest.
• User: "extract claims, methods, benchmarks, and limitations from this paper" -> use paper-digest.
• User: "add this paper to the wiki" -> use paper-digest first if no digest exists, then paper-ingest.

Workflow

1. Identify the source as a PDF URL, local PDF file, extracted full text, abstract, or metadata-only record.
2. Reuse an existing digest when the same canonical paper key is already present.
3. If creating a digest, extract only source-grounded information.
4. Preserve identifiers and technical names exactly: title, DOI, arXiv id, method names, benchmark names, datasets, metrics, theorems, and cited systems.
5. Distinguish full-document evidence from abstract-only or metadata-only evidence.
6. Record uncertainty explicitly instead of filling gaps from general knowledge.
7. Return a structured digest suitable for later paper-ingest.

Schema Contract

Use the current PaperDigest contract as the canonical digest schema. The model-facing output must match the PaperDigestModelOutput shape; runtime records add ids, source identity, canonical key, source kind, content level, and timestamps.

Do not introduce an alternate digest shape unless the code schema has been migrated.

Digest Boundary

Do:

• Summarize the paper's research problem, motivation, approach, contributions, claims, findings, limitations, important terms, and literature-review use.
• Classify the paper as computational empirical, experimental empirical, methodological/instrumentation, theoretical/mathematical, or review/survey when that taxonomy applies.
• Capture specialized fields that matter for later wiki pages, especially methods, benchmarks, findings, formal results, comparators, assumptions, proof ideas, datasets, metrics, and failure modes.
• Preserve extraction caveats so later synthesis can reason about evidence quality.

Do not:

• Create or update wiki pages.
• Decide cross-paper synthesis.
• Merge claims across multiple papers.
• Invent missing bibliographic details.
• Treat abstract-only metadata as full-paper evidence.

Output Quality

The digest should let a later ingest step answer:

• What stable paper page should exist?
• Which existing wiki pages might this paper update?
• What claims, topics, methods, benchmarks, findings, research questions, or formal results are actually justified?
• What should stay uncertain until more source text is available?

When code changes are needed, preserve the schema contract instead of adding a parallel schema.