Zotero Paper Interpretation and Organization
When explicitly invoked with $zotero-paper-review and a paper, the skill defaults to verifying, interpreting, and updating the single matching existing entry in Zotero. If the user specifies only to look up, only to interpret, or only to update certain fields, it works within that scope. When triggered via natural language, it only writes to Zotero when the user explicitly asks to organize it there.
1. Locate Entry and Verify Original Text
Use the zotero:Zotero skill to read entries, child notes, and existing tags. Prefer locating by DOI or entry key, but can also match by title, author, and year. If an entry is ambiguous, first conduct research that can be done independently, then ask the user to select an entry. If no existing entry is found, return the completed interpretation, and only import a new entry upon the user's explicit request.
Open the original paper online and verify the first author, first author's affiliation, journal or conference, publication or public date, paper link, and official code link using the publisher's or proceedings page, preprint page, author's homepage, or official repository. Distinguish between formal publication and preprint, model code and project materials. Record unverifiable literature information as [None], preserving known preprint information. For technical details of the paper, cite the original text's sections, figures, or supplementary materials.
2. Prepare Fields, Notes, and Tags
| Zotero Field | Content to Write |
|---|---|
| Title | Complete or correct the English original title based on the original text, preserving the full subtitle and correct capitalization. |
| Title Translation | Complete Chinese title, using common academic terminology. |
| Short Title | Prefer the method name, abbreviation, or running head short title used in the original text. If the original text has no abbreviation, use Year + First Author's Surname. If the year cannot be verified, leave it as pending. |
| Brief Note | A one-sentence summary of the paper's core contribution; for a survey paper, state the object of organization and the organizational perspective. |
| Abstract Translation | Translate the full original abstract, preserving key terms, numbers, symbols, and resource links. If the original abstract cannot be found, leave this field as pending. |
| Date | Fill in using the official date of the version corresponding to the entry. For officially published versions, use the publication date indicated by the journal or conference proceedings; for preprints, use the first public date. Fill in with YYYY-MM-DD, YYYY-MM, or YYYY based on source precision. If the date cannot be verified, keep the existing value; leave blank fields for later addition. |
Under the entry, organize a child note titled "Short Title · Paper Interpretation". Organize in the following order, using concise Chinese and clickable source links.
- Literature Information – First author, affiliation, journal or conference, year, paper and code links.
- Keywords – If the original text provides keywords, fully reuse their wording, language, and order. If the original text does not provide keywords, extract them based on the paper's topic and main contributions, suggest 4–5.
- Reasoning Input and Modalities – List the data actually received by the model and optional conditions.
- Reasoning Output and Tasks – Distinguish between direct output and downstream applications.
- Model, Training, and Losses – Explain architecture, training method, supervision signal, and main losses.
- Datasets – Distinguish between training, evaluation, and survey collections, briefly describe representative datasets and their uses.
- Limitations – Focus on limitations stated in the original text or supported by evidence.
The Chinese title, short title, brief note, and abstract translation are uniformly stored in the entry's information. For survey papers, first explain the survey's contribution; fill model fields based on applicability. When interpreting representative models covered in the survey, clearly indicate the model's attribution and cite its original paper.
Directly write each keyword from the previous step as a #tag for the entry, using the library's # prefix.
Add status or type tags based on evidence. Add # Open Source if the official model code has been released; otherwise add # Pending Open Source, choosing one of the two. Add # Survey for surveys, and add # Dataset if the main contribution is releasing a dataset or benchmark. Additional tags are considered separately; deduplicate when writing and preserve other user tags.
3. Write and Verify
Edit entry information, child notes, and tags through the Zotero Desktop interface. Use the Zotero local API for read-back verification. If an interpretation note on the same topic is found, update the relevant paragraphs and preserve content added by the user. If multiple candidate notes cannot be distinguished, ask the user to select. If no same-topic note exists, prefer reusing the single blank child note; otherwise, create a new one.
After saving, verify the target entry, the written fields, the parent entry of the child note, and the final tags. Before retrying, first read back the actual state and fill in missing parts to avoid duplication. Finally, briefly state the saved location, tags, and any incomplete fields. If the user chooses to interpret only, return the content in the conversation according to the field table and note structure.