Paper Digest

Paper Digest

Turn a research paper into a structured, critically-read digest. The output is deliberately more than a summary: it separates what the authors claim from what the evidence shows, flags reproducibility risk, and positions the work against prior art. This is the difference between "tell me about this paper" and a digest a researcher can actually rely on for a literature review.

When the input is a PDF (the main case)

Step 1 — Parse the PDF deterministically

Run the bundled script first. Do not eyeball the PDF and start writing — the script extracts verified signals (title, arXiv id, DOI, code links, figure / table / citation counts, section headings, abstract) and a cleaned text dump. Trusting these signals over an unaided read is what keeps the digest from hallucinating result numbers or a nonexistent code release.

python scripts/parse_paper.py PAPER.pdf --out-dir ./out

This writes out/PAPER.meta.json (the signals) and out/PAPER.txt (clean text). Read both. Use meta.json's code_links, arxiv_id, and counts directly in the output — they are verified; your own scan is not.

If the script reports num_pages but near-empty text, the PDF is likely scanned/image-only — tell the user it needs OCR before it can be digested.

Step 2 — Pick the reading lens

Read references/reading_lenses.md and select the lens matching the paper type (empirical / theoretical / survey / systems). Most ML papers are empirical. The lens tells you which sections to scrutinize and what overclaims to watch for — this is where the critical part of the reading comes from.

Step 3 — Read the relevant sections

From PAPER.txt, read the abstract, introduction, method, and experiments/results closely; skim related work and read the limitations/conclusion. Use the section headings in meta.json to navigate. Note the section/table/figure number next to every result you plan to report, so you can attribute it.

Step 4 — Write the digest

Follow references/output_schema.md exactly. Produce both outputs: a Markdown report (the section order is fixed) and the JSON object. Save the JSON to a file (e.g. out/PAPER.digest.json) so it can feed a knowledge base, and show the Markdown report in the response. If a present_files tool is available, present the JSON file.

When the input is an arXiv link or pasted text

No PDF to parse. If it's an arXiv URL, note the id; if the user pasted text, work from that. Skip Step 1, but still apply the lens (Step 2) and the same output contract (Step 4). Be explicit in the report about what you could and could not see (e.g. "results read from pasted abstract only").

Non-negotiable grounding rules

These are restated from the schema because they matter most:

• Never invent a number. Report a metric only if it is in the paper. Missing

→ not reported. Guessing a result is the worst failure this skill can make.

• Attribute every headline result to its Table / Figure / Section.
• Keep the authors' claims and your critique separate. contributions and

evidence report the paper; critique is your independent read.

• Match the user's language for the prose; keep JSON keys in English.

What good looks like

A weak digest restates the abstract. A strong digest does what a sharp reviewer does in five minutes: names the real contribution, points to the one table that carries the paper, identifies the experiment that's missing, and tells you whether you can reproduce it. Aim for that.

Reference files

• scripts/parse_paper.py — deterministic PDF signal + text extraction (Step 1).
• references/reading_lenses.md — per-paper-type scrutiny checklists (Step 2).
• references/output_schema.md — the JSON schema, Markdown template, grounding

rules, and a worked example (Step 4). Read this before writing the digest.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: fdgdws
• Source: fdgdws/paper-digest
• License: MIT

Install and usage instructions live in the source repository linked above.