An Agent Skill that turns research papers into critically-read digests with anti-hallucination PDF parsing and runnable eval suite.
SaferSkills independently audited paper-digest (Agent Skill) and scored it 100/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 0 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 0 flagged
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
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.
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 ./outThis 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.
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.
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.
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.
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").
These are restated from the schema because they matter most:
→ not reported. Guessing a result is the worst failure this skill can make.
contributions andevidence report the paper; critique is your independent read.
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.
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, groundingrules, and a worked example (Step 4). Read this before writing the digest.
~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.