By default, readers have general life experience and comprehension ability, but are unfamiliar with the industry the paper belongs to, and do not know the objects, tools, and ways of working commonly used in that industry. When the user specifies the reader's background, that specification takes precedence.
Start from things readers can directly understand, and explain clearly who did what under what conditions, why they encountered difficulties, which link the authors changed, and how the results changed. Introduce technical names when they help identify the paper's contribution; before readers remember the name, they should already understand the thing it points to.
In the end, readers should be able to explain the paper's ideas in plain language, and also state clearly what the authors actually studied and what contributions they made. When everyday examples carry the explanation, they should clearly correspond to the roles and relationships in the paper, and distinguish hypothetical examples from real experiments.
The whole text should unfold around a clear question, letting readers follow observation, doubt, and checking to understand how the research happened. Concrete objects and actions should continuously carry the explanation; as the explanation deepens, cases or scenarios may be changed, but what they add should be explained. A sense of immediacy comes from details relevant to understanding, not from fabricated psychology or exaggerated conclusions.
When key relationships require readers to assemble them in their minds, prioritize ASCII diagrams to present information flow, before-and-after states, or different dimensions, then read the diagram along with concrete examples. Whether to use diagrams and how many to use is decided by whether they reduce the burden of understanding, not by a fixed story or number of diagrams.
Workflow Routing
| Input | Required materials | Output |
|---|---|---|
| arXiv, PDF, paper link, local paper | ReadingGuide.md、references/template.org、references/paper-map.md |
One Org interpretation note and one paper-map research record |
| Only the paper title | After finding a reliable original text, read the same set of files | Org note and paper-map |
| Explicitly only want a verbal explanation | ReadingGuide.md |
Explain directly without saving files; still satisfy the three goals below |
By default, Org is saved to ~/Context/. The filename follows Denote:{YYYYMMDDTHHMMSS}--paper-{方法名或论文关键词}__paper.org; the timestamp is generated by date +%Y%m%dT%H%M%S . When rewriting an existing note, keep the original file and identifier, verify against the original text, and do not create a duplicate note.
The three goals the interpretation should achieve
- Explain clearly what the paper did. Readers can state the research question, the authors' main contribution, how the method or relationship works, and the scope of application of the key evidence. By the first two sections, they should already know what the paper is studying and what the authors did. Research questions in the paper that cannot be omitted should also be addressed.
- Let readers understand why the original understanding needs to be revised. Start from a concrete problem, explain how it is usually understood, then use the paper's evidence to show where this understanding is insufficient, what the authors added, and how the judgment changed as a result. The change can be a correction, or a supplement, distinction, or narrowing of scope; there is no need to create a dramatic reversal.
- Enable readers to explain the results themselves. If the paper proposes a mechanism that can explain multiple findings, readers should be able to use it to explain at least two different findings and infer what would happen if the relevant conditions were changed. Distinguish conclusions verified by the paper from predictions made based on the mechanism. When the paper has no unified explanation, truthfully explain the relationships among the findings and the unresolved problems.
Check the three goals separately. Explaining smoothly, being accurate, and being able to explain the results each have their own requirements and cannot substitute for one another.
Places where mistakes are easy to make
- The example may be clear, but the paper may still not be clear. When readers cannot state the research question or the main contribution, return to the original text and reorganize the explanation. Fixing sentences alone cannot supply the missing content.
- A statement that is obvious within the field may still be unfamiliar to outsiders. The example itself must also match the reader's starting point. Replacing terms with another set of industry words, or returning to a technical overview immediately after the example, makes readers take on the work of translation again; the concrete actions should continue to explain the relationships that follow.
- Do not force one example to carry the entire paper. When one case cannot illustrate all the research questions, you may switch examples while keeping the connection between the question and the explanation.
- State the contribution concretely. Write out what the authors' method, theory, evaluation, resource, or system specifically adds. An evaluation finding a difference does not mean the cause of the difference has been explained.
- Highlight the main points, but also retain important branches. The paper-map records each necessary line of research and its evidence; the main text explains how they jointly answer the paper's question, and the entire paper must not be reduced to a single central experiment.
- Select evidence by its informational role. Prioritize explaining the most representative results. Supplementary results should provide different information, limit the conclusions, or explain exceptions; do not allocate space evenly by number of experiments, and do not insist that the whole text present only one kind of result.
- Numbers need context. Explain what the experiment is asking, what was held fixed, and what it is compared with, then state which judgment the result changes. Breaking the numbers into short paragraphs, or writing them all out in words, cannot replace the explanation.
- Sections may build progressively or stand in parallel. Reordering is used to check whether the connections are clear. Two independent experiments may be swapped in order, but you should explain what each supports and why both are needed.
- Checking tools can only check format and records. They cannot prove that the article is easy to understand. The comprehension check is performed by a new context that did not participate in writing; the evaluator reads only the final notes and, according to
ReadingGuide.mdpoints out places that require rereading, added background, or filling in the picture independently; when there are figures, verify the meaning of the figures, and do not judge readability merely by whether it can be retold.
Examples
Methods paper: Use a small task to explain where the old approach gets stuck, then explain how the new method processes the input, which key steps it goes through, and what results it produces. The central experiment explains the comparison conditions and the differences in actions, and the ending states at which step the improvement occurs and which cases remain unverified.
Evaluation paper: Two systems perform similarly on the old metric, and the new evaluation reveals differences that were not previously distinguished. Explain what the new evaluation measures, what the representative comparison shows, and which judgments these differences can support. Do not write measurement results directly as causal conclusions about reality.
Saving and checking
For the specific writing method, see ReadingGuide.md. The main text uses references/template.org, and the research and trade-off records use references/paper-map.md. New records use format version 3; the validation script is compatible with older versions 1 and 2, and old records do not need to be rewritten in bulk for this upgrade.
After writing, run:
bun {skill_dir}/scripts/validate_note.ts /absolute/path/to/note.org \
--map /absolute/path/to/paper-map.md
Use real Emacs to read back the identifier, file name, directory index, consult-notes, and org-lint. Then complete the independent reading check according to the guide, recording the four results: full-text retelling, research question and contribution, change in understanding, and interpretation of results. After all pass, run the script again to confirm that the final files are consistent with the check records.
The validation script accepts records for which the reading check result has not yet been filled in for diagnostic purposes, but it will return a failure; you cannot first write PASS and then ask someone to check. When a validator or independent evaluation is unavailable, clearly report which item has not yet been completed.