| name | paper-digest |
| description | Read and analyze AI research papers from a PDF or URL and produce a structured markdown digest grounded in the paper. Use when the user wants a paper summary, paper digest, benchmark or dataset breakdown, method breakdown, survey summary, position-paper analysis, experiment summary, or a structured reading note with faithful extraction of contributions, experiments, tables, and figures. |
Paper Digest
Purpose
Read an AI research paper from a PDF or URL, identify the paper type, and produce a structured markdown digest grounded strictly in the paper.
Workflow
- Read the paper fully enough to identify its title, main contribution, paper type, structure, and evidence.
- Classify the paper using references/paper-types.md.
- Extract information using references/extraction-rules.md.
- Build the digest using references/output-schema.md.
- Preserve original contribution wording when the workflow asks for direct extraction.
- Reproduce table contents in Markdown when they are readable from the source. If a figure cannot be reproduced faithfully as a table, describe it faithfully and specifically instead of inventing details.
- Keep every summary, analysis, and interpretation strictly grounded in the paper.
Output Rules
- Return Markdown only.
- Always include the common sections:
Title
One-line summary
Paper type & keywords
Author-stated contributions
- Add the appropriate paper-type-specific sections only after classifying the paper.
- Do not add personal opinions or outside evaluation unless the user explicitly asks for them.
- Do not invent missing statistics, dataset properties, model settings, citations, figure values, or claims.
- If a requested item is missing or unclear in the paper, say so explicitly.
Input Handling
- Accept either:
- a local PDF path
- a paper URL
- When both PDF text and layout cues matter, prefer the paper itself over secondary metadata.
- If the input URL points to a metadata or abstract page, try to access the full paper text before producing a full digest.
- If only abstract-level or otherwise partial content is accessible, produce a limited digest and explicitly state which sections cannot be completed reliably.
- If the source is long or dense, extract the digest section by section rather than relying on a shallow skim.
Paper-Type Routing
- For benchmark or dataset papers, emphasize dataset structure, construction, statistics, and experimental usage.
- For modelling papers, emphasize method logic, components, design choices, assumptions, and empirical results.
- For survey papers, emphasize organizational logic, section-wise takeaways, and influential cited work.
- For position papers, emphasize the core position, argument flow, alternative positions, and refutation structure.
Fidelity Rules
- Preserve author-stated contributions as directly as possible when extracting them from the paper.
- Preserve original tables in Markdown when readable.
- Describe figures faithfully when direct tabular reconstruction is not realistic.
- Distinguish direct extraction from brief synthesis where needed, but keep both grounded in the paper text.