Paper X-Ray

A paper is months of work compressed into a dozen pages. The method section is in teaching order, and the introduction tells a story assembled after the fact. Most of the judgment calls that decided the outcome never make it to the page: what the authors saw, where they got stuck, how the first version failed, which oddly specific hyperparameter is guarding against which failure.

Paper X-Ray recovers that. Give it a paper, or a topic, and it will:

• Reconstruct the bet and the detours. The concrete scene where prior methods break, what gave the authors confidence, and what the strength of their wording reveals about where they are sure and where they are bluffing.
• Write from the lab bench. After every mechanism, a passage on how the first version fails, the cheapest diagnostic a lead would run, and the patch with its cost.
• Make the math hand-checkable. Every symbol gets its shape and meaning; every key formula gets a small example you can verify on paper.
• Read skeptically. Numbers are checked against noise and against each other; what the paper does not report is marked as unreported.
• Grade its evidence. Every inference is tagged as stated in the paper, inferred, or guessed. Bold guesses are fine; guesses dressed as facts are not.

The voice is Hinton-plain and the exposition is 3Blue1Brown-style: see it first, then compute it. Output is either a long-form Markdown document or a single-file interactive HTML page.

Each is one HTML file. Download and open it in a browser; no build, no server. The demos are written in Chinese.

Crops from the LLM demo. Every figure is drawn in one figure language: color marks a role rather than decorating, a box holds a name and nothing else, wires run only horizontally and vertically, and no figure carries a title or a legend.

Left to right, top to bottom: drag the compute budget and watch the bottom of the isoFLOP curve move; step Kaplan's 0.835 down to Chinchilla's 0.497 one correction at a time; a byte-level BPE trained on the page itself, re-segmenting as you drag the merge count; Llama 3 8B's eight billion parameters taken apart layer by layer; the bolt-on and native routes to multimodality (dark theme); GRPO advantages as you change how many of 16 answers are correct (dark theme).

The overview opens into detail. Each step lists what three real training runs chose and the small experiment that decision rested on:

A slider always drives a figure that changes, never just a number in a table. Here, tokens per frame decide how much video fits in a million-token context:

The HTML ships with its own annotation layer; readers install nothing. Turn on ✎, select text, and highlight it, write a note, or mark it as lost, too long, or got it. Notes sit in the right margin, tied back to the text by a thin line.

⤓ writes the annotations back into the HTML file itself. The next time the skill runs it reads them: a concept marked lost is explained again from a different angle, a kind of passage repeatedly marked too long gets shorter by default, and a concept marked got it is mentioned in one line from then on. The reader profile and the log live in a local user/ folder that never enters the repository.

One-line prompt: paste this into your agent.

Clone https://github.com/Wang-auspicious/paper-xray into a folder named paper-xray in my agent skills directory, then confirm the skill is registered and callable.

Git clone (recommended; brings the references and the figure kit along):

git clone https://github.com/Wang-auspicious/paper-xray.git ~/.claude/skills/paper-xray

On Windows the skills directory is C:\Users\<you>\.claude\skills\paper-xray. Copying only SKILL.md works for Markdown output, but HTML output needs the page skeleton and figure kit in references/.

/paper-xray D:/papers/rope.pdf
/paper-xray 2104.09864 html
/paper-xray the full training pipeline of a foundation model html

1. Provide the material: a local PDF, an arXiv identifier or link, pasted text, or a topic. Attach public code when there is some: normalization, warmup and data filtering that only exist in the code are sometimes where the results actually come from.
2. Pick a format: md for the long-form text, html for the interactive page. If you name neither, it asks once.
3. Let it read: it reads the body, appendices, footnotes and captions, then writes section by section without compressing to fit a single response. For a topic too big for one document, it first proposes a map of how to split it.

On first use it asks a short questionnaire about your math and hands-on background and adjusts the writing to match.

paper-xray/
├── SKILL.md                       # Chinese specification: workflow, writing rules, topic mode, reader profile
├── VERSION
├── references/
│   ├── SKILL.en.md                # English specification
│   ├── html.md                    # HTML delivery spec: page, figure language, interaction, annotation, self-checks
│   ├── reader-kit.html            # Page skeleton: layout, contents, themes, annotation layer, XRAY.lint
│   └── figure-kit.html            # Figure components: f.row / f.zoom / FK.plot / FK.tokens / FK.heat / FK.iso / FK.lint
├── demos/
│   ├── llm-pretraining-xray/      # Topic-mode demo with its body, figure code, tokenizer trainer and build script
│   ├── preln-xray/                # Pre-LN paper demo with the minimal reimplementation
│   └── perovskite-ion-xray/       # Perovskite ion-migration demo with the hand-calculation script
├── assets/                        # Images for this README
├── README.md
├── README.zh-CN.md
└── LICENSE

If this saved you a reread, a star helps other people find it.

MIT. See LICENSE.