A structured, fast read of a research paper that surfaces the claims, methods, evidence, and limitations — not a generic "this paper is about X" summary.
When to use
- An arxiv preprint, journal paper, conference proceedings PDF
- A technical white paper from a company (model card, system card)
- A long internal research doc the user wants triaged
Skip for: short posts (blog, news, social), commercial reports without a methods section (those need a different lens), and anything where the user actually has time to read it themselves — the value of this skill is triage, not replacement.
Workflow (a five-pass read)
Title + abstract + figures. That's pass 1. Don't open the introduction yet. From these three you should be able to state: what's the claim, what's the evidence, what's the comparison. If you can't, the abstract is bad or the claim is being hidden — flag it.
Methods section: identify the experiment shape. Is it a benchmark comparison (model A vs B on dataset X)? A theory paper with a proof? A new technique introduced + tested on N tasks? An ablation study? Each shape has a different way of being load-bearing.
Find the central table or figure. Most papers have ONE table or figure that carries the claim. Open the methods section, find the bigest table referenced in the abstract, and study it. Note: which baseline they compare against, how they chose the baseline, what dataset they tested on, what's the magnitude of the effect.
The limitations / threats-to-validity / failure-modes section. Authors write these sections carefully — they list the caveats that, if discovered by a reader otherwise, would torpedo credibility. Read this section like an adversary. What did the authors choose to flag?
The "related work" placement check. Where does this paper position itself? If they cite 30 papers but the most-similar prior work is mentioned briefly and dismissed, that's a smell — either the authors don't know the field or they're hand-waving past a precedent.
Output shape
**Paper**: <title> — <author 1>, <author 2>, ... (<venue>, <year>)
**Claim**: <one sentence — what's new and what improves>
**Method** (sketch): <2–3 sentences. Not the title of the method; the actual mechanism>
**Evidence**: <main table/figure result with numbers — N improvement on M benchmark, statistical sig if reported>
**Comparison**: <what baseline; how the comparison was fair or not>
**Limitations** (per authors): <2–3 bullets, verbatim from the paper>
**My skepticisms** (per skim): <2–3 bullets — what jumped out as load-bearing but undertested>
**Read-the-rest** signal: <Y/N + why. If Y, what section>
The "my skepticisms" beat is the load-bearing one — it's why the user delegated the read in the first place.
Definition of done
- Claim stated in one sentence
- One concrete number (effect size, accuracy %, etc.)
- Comparison baseline named
- At least 2 limitations identified
- At least 2 skim-level skepticisms surfaced
- A clear "yes, read the rest" or "no, this is enough" verdict
Gotchas
Don't paraphrase the abstract. That's not a skim — that's a copy. The abstract is what THEY want you to take away; the skim is what you ACTUALLY learned. Disagree where appropriate.
Numbers without baselines are noise. "Achieves 87% accuracy" is meaningless without "vs 82% prior SOTA" or "vs random = 50%."
"State-of-the-art" without specifying the leaderboard / split is a hand-wave. Papers claim SOTA against cherry-picked baselines. Note which leaderboard / which split.
Beware ablation absence. A paper introducing 5 techniques but only ablating 1 is hiding 4 unproven decisions. Flag it.
Replicability check. Is code released? Are weights released? Are evaluation prompts in the appendix? Each absent piece reduces credibility by a recognizable amount.
Be calibrated. Sometimes the paper is genuinely good. Don't manufacture skepticisms to look sharp. Calibrated "this looks solid; one note on the choice of dataset" beats theatrical critique.
Watch for fabricated citations in AI-generated summaries. If your skim is going to be used downstream, double-check the year/author of any cite the user might quote. Real cite patterns over-specify (volume, page, DOI); fabricated ones under-specify.