ljg-xray-paper

LJG-Xray-Paper: Paper Interpretation

You need to do two things, only two:

1. What the paper says: problem → perspective → result
2. What it means to me: cognitive cards (ASCII art visually presenting insights)

Everything else serves these two things.

Constraints

L0: General constraints

Org-mode syntax

• Bold uses *bold*(single asterisk), forbidden **bold**
• Heading levels start from * , no skipping levels

ASCII Art

All diagrams, topologies, and cards must be drawn using pure ASCII characters.

Allowed character set:+ - | / \ > < v ^ * = ~ . : # [ ] ( ) _ , ; ! ' " and spaces.

All Unicode drawing symbols are forbidden, including but not limited to: ─ │ ┌ ┐ └ ┘ ├ ┤ ┬ ┴ ┼ ═ ║ ╔ ╗ ╚ ╝ ╠ ╣ ╦ ╩ ╬ ▼ ▲ ► ◄ → ← ↑ ↓ ● ○ ■ □ ◆ ◇

Exception: skills whose output target is an HTML file (browser rendering) are not subject to this restriction.

Denote file specification

• Timestamp retrieval:date +%Y%m%dT%H%M%S
• Human-readable time retrieval:date "+%Y-%m-%d %a %H:%M"
• Filename format:{时间戳}--{标题关键词}__{标签}.org
• Output directory:~/Documents/notes/
• Exception for visual outputs (HTML/PNG):~/Downloads/ or /tmp/

Org file header

#+title:      {标题}
#+date:       [{YYYY-MM-DD Day HH:MM}]
#+filetags:   :{标签}:
#+identifier: {YYYYMMDDTHHMMSS}

Completion action

After the file is written, report the file path to the user.

L1: Cognitive constraints

Cognitive baseline loading

Before executing the analysis, read the following files to establish a cognitive baseline:

1. ~/Documents/know/soul.md — worldview, thinking paradigms, core beliefs
2. ~/Documents/know/memory.md — long-term memory, knowledge connections

Honesty principle

• delta ≈ 0 is a normal result, do not force it
• If there is no collision, do not fabricate a card
• Information that cannot be found is marked "insufficient information", do not fabricate
• If it cannot be compressed into one sentence = not yet thought through, keep thinking, do not fudge it

Cognitive collision cards

Card = ASCII art, visual first.

The standard for a good card: cover the text and look only at the lines, and you can still feel the structural relationships (branching, convergence, hierarchy, contrast, tension).

Counterexample—a text list disguised as a card:

+------------------+
| 论文说: X        |
| 我原来想: Y      |
| 现在变成: Z      |
+------------------+

This is not a card, it is text with a frame. Structural relationships must be expressed through spatial layout, not declared with labels.

Each card anchors to a specific scenario: a way of thinking, a decision scenario, or a change in a cognitive blind spot. Below the card, attach a quote-level insight—one that can stand alone outside its context.

Execution steps

Step 1: Receive the paper and preprocess the input

User input	Conversion rule
2601.01290 or arxiv:2601.01290	→ https://arxiv.org/html/2601.01290
https://arxiv.org/abs/... or .../pdf/...	→ replace with /html/
https://arxiv.org/html/...	use directly
Other URLs or PDF paths	process according to the original logic

arxiv papers are always converted to /html/ format; the HTML version can be fetched in full via WebFetch.

Extract the title and authors, and fill them into the report header.

Step 2: Load the cognitive frame of reference

Execute cognitive baseline loading (see constraint L1).

Step 3: What the paper says

Like telling a smart friend at the dinner table "what this paper did"—three sentences, plain language, no paper clichés:

• Problem: What are the authors trying to solve?
• Perspective: What method/angle do they use?
• Result: What did we get? Naturally distinguish "experimentally proven" from "author's speculation." If a speculative conclusion is interesting, turn it into an "open question" collision card in Step 4.

Zero-jargon rule: Every technical concept must immediately land on something the reader has seen. If you don't give an example, you're not allowed to use that word.

Bad: "Co-occurrence probability depends only on distance in the latent space (translational symmetry)" Good: "January and February often appear in the same passage, while January and July rarely do—the closer the distance, the more often they appear together"

The reasoning must grow out of the scenario, not be stuck on as a label.

Scenario-ize the load-bearing concepts: Every paper has 2-3 "load-bearing concepts"—remove them and the paper's argument collapses. These concepts cannot be glossed over with a parenthetical note ("gamma: conditional entropy decay exponent"). A parenthetical note is a hint for people who already understand, not an explanation. For each load-bearing concept:

1. First find a scenario the reader has personally experienced (group chat logs, waiting in line, finding your way...)
2. Unfold it step by step within the scenario—at least 3 levels of progression, so the reader can "feel" this concept changing in something familiar
3. Only at the end attach the technical name. At this point the name is a label, not a definition

Test: remove the technical name, and the reader still knows what you're talking about = successfully grounded. Only a parenthetical note remains = failure.

The relationships between concepts should likewise be scenario-ized. Not "alpha is the ratio of gamma to beta," but "the cost-effectiveness of distant information: how many points is it worth / how hard is it to reach = how steep the scaling is."

Ordinary terms only need a one-sentence analogy, no scenario unfolding required. Distinguishing load-bearing from non-load-bearing is the key judgment.

Finally compress it into one plain-language sentence—like you're in an elevator telling a friend who knows nothing about this field, "This paper is basically saying..." No jargon allowed, no more than one sentence. If you can't compress it = you haven't thought it through yet, keep thinking.

Include a napkin sketch (ASCII) that draws the core mechanism.

Step 4: What it means for me

After reading this paper, what do I take away?

First answer in one sentence: what might this paper's ideas change about me? Point to a specific thinking habit, decision pattern, or cognitive blind spot. Not "helps me understand X better"—that's an information gain, not a change. If you can't find a change, honestly write "delta ≈ 0."

soul.md + memory.md provide the background of "who I am," but the insights are not limited to already-recorded entries. Any insight that can improve the quality of my decisions deserves a card:

• A thinking tool I can use directly (that I didn't have before)
• New evidence that changes some judgment
• An angle or blind spot I hadn't thought of
• A supplement or correction to existing knowledge

For each insightful point, generate a cognition card (ASCII art). See constraint L1 for card quality standards.

The card should intuitively show: how this insight changes some way I think, a decision habit, or opens up a blind spot. One glance and you get where the insight is.

Good card examples for this skill—bifurcation type:

  Soul: "做了才懂"
        |
        |  但如果结构藏得太深...
        v
  +------------------+
  | 反馈能暴露结构?   |
  +---+----------+---+
      |          |
     YES         NO
      |          |
   生成优先    降秩优先
  (原路径)   (先看再做)
      |          |
      |     CTA: 0% vs 94%
      v          v
   Soul 成立   Soul 需加边界

Threshold type (not linear, there's a turning point):

  我以为: 投入越多，产出越多（线性）

  投入 ----+----+----+----+----+---->
           |    |    |    |    |
  产出     .    .    .    .    .   线性预期
           .    .    .   /
  实际     .    .    . _/         <-- 阈值
           .    .    ./
           .____.___.'............   无效区
           |
       "原来这段全是白费的"
           |
           v
  决策改变: 不追问"做了多少"
            要追问"过没过线"

Collisions have many shapes—bifurcation, tension, threshold, gap, flip. Choose the spatial structure that matches the collision itself; don't compress all collisions into the same kind of diagram.

If the paper's speculative viewpoints are interesting, also turn them into collision cards—labeled "open questions," with only a direction and no conclusion. Write a sharp question, not a gentle checkbox. If it's worth pursuing, it can be pursued; if it's not worth pursuing, that's clear at a glance.

Step 5: Generate the Org report

1. Obtain the timestamp according to the Denote file specification (see constraint L0)
2. Filename:{时间戳}--paper-{简短标题}__read.org
3. Read references/template.org to obtain the report structure, and fill it in according to the template
4. Write to ~/Documents/notes/{文件名}

Output quality standards

• Only two parts: what the paper says + what it means to me. Add nothing else