agentleFS
tyler-voidtyler-void/truetake/skills/paper-reading/SKILL.md

paper-reading

Skill1 starsChanged 6 days agoMITView on GitHub

Read a research paper the way a researcher does before explaining it, and write down what an explainer needs - the method in execution order, every tensor and size, every number with its table or figure, what was released, and what the paper leaves undefined. Use before planning any explainer video, figure or summary of a paper, or when asked to read, digest or extract a paper.

22 other projects use a version of this.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

4 rules. 0 are common, 3 are distinctive, and 1 is very specific to this project.

  • 0

    Common

    Most of the 12 edited versions we read keep these.

  • 3

    Distinctive

    This author's own choices. Worth a look.

  • 1

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Read the whole paper three times, each time for something different.

    line 17
  2. Distinctive:

    How it works, in the order the code runs: Read the method section with the framework figure open.

    line 22
  3. Distinctive:

    Find what was released: before planning: repository, licence, weights, data, and what is *not* there.

    line 64
  4. Very specific:

    When the reading is done, tell the user the claim, the modules in order, the gaps and their assumptions, and what was and was not released, …

    line 76

“Common” is measured: kept by at least 6 in 10 of the 12 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Reading a paper before explaining it

An explainer is only as true as the reading behind it. This skill turns a paper into paper.json and a short method.md, which every later step (scene plan, captions, numbers on screen) cites. Nothing goes on screen that cannot be traced to a line in one of them.

The three passes

Read the whole paper three times, each time for something different. Do not skip to the method.

  1. What it claims (10 minutes). Title, abstract, introduction's contribution list, every figure and its caption, conclusion. Write the claim in one sentence, the problem in one sentence, and the contributions as a list. Note which contribution is the method (what a reader could re-implement) and which are data or benchmarks.
  2. How it works, in the order the code runs. Read the method section with the framework figure open. Write the modules in execution order: for each, its input, its output, its size at each step, and the equation or algorithm that defines it. Follow one sample through; if the paper never says a size, compute it from the architecture and mark it derived.
  3. The details a re-implementer would trip on. Hyperparameters (with the sentence that states them), thresholds, training versus inference differences, pre- and post-processing (resize, normalisation, NMS), every number in the tables you may show, and every place the paper is silent.

Tables and figures are often images in the PDF: render the page (pdftoppm -r 200) and read them; never guess a number from text that did not extract.

What to write

paper.json, in this shape (keys may be empty, never invented):

{
  "paper": {"title": "", "authors": [], "venue": "", "year": 0, "doi": "", "arxiv": "", "license": ""},
  "claim": "", "problem": "",
  "contributions": [{"kind": "method|data|benchmark|analysis", "text": ""}],
  "modules": [{"name": "", "order": 1, "input": "", "output": "", "sizes": [""], "defined_by": "Eq. 3 / Alg. 1 / Sec. III-B",
               "parallel_with": []}],
  "equations": [{"id": "Eq. 1", "tex": "", "symbols": {"x": "meaning"}, "where": "Sec."}],
  "hyperparameters": [{"name": "", "value": "", "quote": "", "where": ""}],
  "numbers": [{"value": "", "meaning": "", "where": "Table III, row SiamFC, column noDET"}],
  "released": {"code": "", "weights": "", "data": "", "licenses": ""},
  "gaps": [{"what": "", "why_it_matters": "", "our_assumption": ""}],
  "figures_to_redraw": [{"figure": "Fig. 7", "why": "", "license_note": ""}]
}

method.md: the method in plain words, one short paragraph per module, in execution order, each ending with the source it came from. This is what a reviewer (the paper's own field) reads to check the explainer.

Rules

  • Quote, do not paraphrase, every hyperparameter and threshold, with its section. "τt and τd are set to 0.9" beats "high thresholds".
  • Every number carries its location: table, row, column; or figure and axis. Numbers measured by us are never presented as the paper's, and the paper's are never presented as ours.
  • Separate the paper's method from what it uses. A tracking paper that plugs in Faster R-CNN does not own Faster R-CNN; say which parts are the contribution.
  • Find what was released before planning: repository, licence, weights, data, and what is not there. A video that needs the authors' weights cannot be made from a paper that did not release them; plan with what exists and label substitutes.
  • List the gaps. Anything undefined that an explainer must decide (a score's normalisation, an unstated input size) goes in gaps with the assumption chosen. The explainer states that assumption on screen.
  • Licences of the paper itself: the publisher's version is usually all rights reserved; the arXiv version may carry a Creative Commons licence. Redraw figures from your own real data either way, and cite the original.
  • Cross-check against code wherever code exists (sign conventions, transposes, inverse matrices, the order of steps), and record what was checked.

Hand-off

When the reading is done, tell the user the claim, the modules in order, the gaps and their assumptions, and what was and was not released, in a few lines. The explainer plan starts from paper.json, not from the PDF.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.