--- name: "reading_papers" description: "Read academic papers efficiently: run the three-pass method (Keshav), pull the claim from the abstract and conclusion first, evaluate methods sections honestly, and build a compounding paper-note pipeline. Triggers when the user needs to read, summarize, or critically assess a research paper." --- # Reading Papers ## Purpose Read academic papers without drowning in them. The core method is S. Keshav's three-pass system ("How to Read a Paper", ACM SIGCOMM Computer Communication Review 37(3), 2007): each pass has a fixed goal and budget, and you only spend the next hour when the paper earned it. Most papers get a 10-minute pass and a decision. Only the ones that matter get the deep read. ## Workflow ### 1. Decide if the paper deserves your time (Pass 1, 5-10 min) Read in this order: title, abstract, introduction (first paragraph or two), section and subsection headings, conclusion, references (glance; note any you recognize). Skip the body entirely. Answer the five Cs: 1. **Category:** what kind of paper is it (measurement, analysis, method, survey, position, benchmark)? 2. **Context:** what line of work does it sit in, and which papers does it build on? 3. **Correctness:** do the assumptions look defensible at a glance? 4. **Contributions:** what does it claim is new, one sentence each? 5. **Clarity:** is it well written, or will it fight you? Then gate: outside your scope, log it as skimmed and stop. Adjacent to your work, go to Pass 2. Core to what you're doing, plan for Pass 3. ### 2. Grasp the content (Pass 2, up to 1 hour) Read the whole paper but skip proofs, derivations, and implementation detail on purpose. Study every figure and table first, then read the text around them: check axes labels, error bars, sample sizes, and whether significance is actually reported. Mark references that look central and go chase them later. Take margin notes as you go, especially questions. End of Pass 2 output: a short paragraph summarizing the paper's argument, and a list of open questions. You should be able to explain the main contribution to a colleague in a couple of minutes with evidence behind it. ### 3. Re-implement mentally (Pass 3, 1-5 hours) Reserved for papers you need to fully own, e.g. to review them, build on them, or refute them. Reconstruct the work from the authors' assumptions: trace each equation and each number in the key tables, ask what would happen if an assumption changed, and look for what they left out (failure modes, hyperparameter sensitivity, negative results they buried). Ask what a skeptical reviewer would ask. End of Pass 3 output: you could re-derive the result from memory and you have found at least one real weakness. ### 4. Evaluate the methods honestly For empirical papers, run the methods checklist in `references/methods-checklist.md`. The short version: is the study reproducible from the description, is the sample big enough for the claimed effect, are the stats appropriate for the data type, was multiple testing corrected, are effect sizes reported alongside p-values, and is statistical significance being confused with practical importance? A paper can be clearly written and still have shaky evidence. Say so plainly. ### 5. File the note so it compounds Write the note immediately after reading, same session, using `assets/paper-note-template.md`. Keep a running note folder (one note per paper, plus a simple index by topic). When a note references another paper, link them; over time the links become a map of the field. This is what turns reading into a pipeline instead of a pile. ## Output Contract Depending on the pass run, return: - Pass 1: the five Cs (a few lines each) + the depth decision. - Pass 2: a paragraph summary + open questions + references to chase. - Pass 3: the summary plus the reconstructed argument, identified assumptions, and at least one weakness. - Note file: filled `assets/paper-note-template.md` when the user wants the note kept. Always say which pass you ran and why. If the paper's figures are sloppy (unlabeled axes, missing error bars), flag that as a quality signal, not just a nitpick. ## Operating Rules 1. Never read a paper front-to-back as the first approach. Pass 1 first, every time. 2. Most papers die at Pass 1. That is the method working, not failing. 3. Every factual claim in a summary must come from the paper itself. Unverifiable interpretations get labeled as opinion. 4. Figures get more scrutiny than prose. If the numbers in the table don't match the story in the text, trust neither until you check. 5. Write the note right after reading. Waiting even a day measurably degrades recall. 6. The pass is the unit of effort, not the clock. An expert may not need the full Pass 3; scale the budget to your familiarity, not to a stopwatch.