AGENT SKILL

paper-deep-read

Explain one difficult paper at the student's level — its objective, its method step by step, its mathematics, its hidden assumptions and what is reproducible. Use when the student asks what a paper did, how a method works, why it is hard, or says they do not understand an article.

License 许可
MIT
所属仓库
LNdme/SUniv
最近更新
2026-11-12
Stars
★ 1
Forks
0

01 前言 Front Matter

name
paper-deep-read
description
Explain one difficult paper at the student's level — its objective, its method step by step, its mathematics, its hidden assumptions and what is reproducible. Use when the student asks what a paper did, how a method works, why it is hard, or says they do not understand an article.

03 SKILL.md 正文

paper-deep-read — make one hard paper genuinely understood

The student is not asking for a summary. They are asking to be brought to the point where they could explain the paper to their supervisor, or build on it. Those are different outputs, and only one of them is worth their time.

Get the text first

You cannot deep-read a paper you have not read. Check grounding before starting:

suniv-search --doi <doi>                      # is there an open-access copy?
suniv-lib search "<title words>"              # does the student already hold it?
suniv-lib sections --file <path>              # the structure, with page numbers
suniv-lib text --file <path> --page 4         # one page at a time

If only the abstract exists, say so and offer what an abstract can support: the question, the claimed contribution, the venue. Do not walk through a method you have not seen. If nothing but metadata exists, this skill does not apply yet — help the student obtain the text.

Calibrate to the person

Ask what they already have, or infer it from how they asked. A first-year master's student meeting variational inference and a doctoral student who knows it but not this variant need different explanations, and giving the wrong one wastes the session. When unsure, start one level below where you think they are and move fast — being briefly over-explained costs seconds, being lost costs the afternoon.

The order that works

  1. The problem, before the solution. What was broken or unknown, and why did existing approaches not settle it? A method makes no sense until the student feels the difficulty it answers.
  2. The idea in one sentence, in plain language, no notation.
  3. The method, step by step. What goes in, what happens to it, what comes out. Name each step's purpose before its mechanics.
  4. The mathematics, once the shape is clear. Define every symbol on first use. Say what each equation does — this term penalises that, this expectation is over that distribution. An equation restated in words is not an explanation; an equation whose role is explained is.
  5. The evaluation. What is compared to what, on which data, and does the comparison actually support the claim? This is where papers are weakest and where a student learns to read critically.
  6. Assumptions and limits. What must be true for this to work? What do the authors concede, and what do they pass over quietly? Distinguish the two — only the first is on the record.
  7. Reproducibility. Is there code, data, hyperparameters, compute? What would the student need to run it themselves?

Attribute everything

Every claim about the paper points to where it is: a section heading, a page, an equation number. "§3.2 defines attention as…" lets the student check you and find it again. "The paper says…" does not, and is where a misreading becomes theirs.

When you are inferring rather than reporting — reading between the lines, connecting to something the paper does not say — mark it as your reading. That distinction is the difference between teaching and misleading.

Close usefully

  • what the student should now be able to do that they could not before;
  • the two or three papers this one rests on, if they need the foundation — retrieve them with suniv-search, do not name them from memory;
  • the honest verdict on quality: what this paper establishes, and what it only suggests.

Then offer to file the notes into Zotero against the item, so the reading survives the conversation.

04 同仓库更多 Skills

LNdme
LNdme/SUniv

use-shared-credential

When you need to call a service the org has a SHARED credential for (a SERP/search key, a paid-API key, a data-vendor feed, a Git remote) — and the platform has told you that credential is available to this conversation — make the call BY PROXY through the credential broker. You never see the secret; the core stamps it onto the outbound request for you.

★ 1
LNdme
LNdme/SUniv

taste-skill

Design taste and process for anything a person will look at in a browser — landing page, dashboard, prototype, deck. The default design driver for a browsable artifact, with an anti-slop playbook for the visual language.

★ 1
LNdme
LNdme/SUniv

publish

Publish a long-lived internal web app, site, or dashboard from the agent computer. Scope-bound (only the owner's scope, plus whoever you share it with, can reach it), immutable versions with rollback, a stable friendly link. No public URL.

★ 1
LNdme
LNdme/SUniv

popular-web-designs

54 real-world design systems (Stripe, Linear, Vercel, Notion, Apple…) as ready-to-paste HTML/CSS — exact colors, typography, components, and spacing for styling a page after a known brand.

★ 1
LNdme
LNdme/SUniv

morning-digest

Assemble and deliver one short morning digest for the person who owns the cron — what changed overnight in the sources they already connected, and what it means for their day.

★ 1
LNdme
LNdme/SUniv

memory

Deliberately search, add to, or curate your long-term memory with the `memory` tool — beyond the automatic recall/capture every turn already does. Use when asked who someone is or what they're working on, when you learn a durable fact worth keeping now, or when your notebook holds stale or wrong entries you should clean up.

★ 1