| name | paper-discussion |
| description | Use when the user asks substantive follow-up questions about a specific paper or asks to discuss, explain, critique, or connect a paper to their research. Maintain one evolving local note per paper in data/notes/ARXIV_ID.md, read local notes and source before analysis, and integrate feedback-memory during sustained discussion. |
Paper Discussion
Use this skill for sustained or explicit discussion of one paper. Do not use it for ordinary daily ranking unless the conversation shifts into discussing a specific paper.
Trigger Rules
Use this skill when:
- The user asks a substantive question about a specific arXiv paper.
- The user asks to explain, discuss, critique, interpret, or connect a paper to their work.
- The user continues asking follow-up questions about the same paper.
- A
data/notes/ARXIV_ID.md note exists and the user asks about that paper.
Do not create a note for feedback-only turns. If the user only gives feedback such as +1, interesting, not useful, or a reusable preference signal, invoke feedback-memory; update the paper note only if it already exists.
Local Order Of Operations
For substantive paper discussion:
- Resolve the arXiv ID.
- Read
data/notes/ARXIV_ID.md first if it exists.
- Locate local metadata/reports to identify title, source date, and prior recommendation context.
- Check for existing local paper material under
data/raw/arxiv/*/sources/ARXIV_ID/.
- If local source/PDF/text is missing, run:
.conda/arxiv-daily/bin/arxiv-daily download ARXIV_ID
- Re-check the local source folder.
- If no local paper artifact can be obtained, report that grounded discussion is blocked and avoid deep analysis from memory or abstract-only context.
Use local source as the source of truth. Use network only for missing source acquisition, resolving absent metadata, or version/publication checks explicitly requested by the user.
Source Reading
Prefer targeted local inspection:
- Existing note first.
- Abstract, introduction, result/theorem statements, method sections, discussion, limitations.
.bbl / .bib only if needed for context.
Do not read whole large TeX trees unless necessary. Use rg and snippets.
Paper Notes
Maintain one evolving note per paper:
data/notes/ARXIV_ID.md
The note should be compact but self-contained. Use flexible sections appropriate to the paper, for example:
- Core Idea
- Setup
- Main Result
- Mechanism
- Technical Stack
- Weaknesses And Caveats
- Takeaway
- Promising Future Ideas
Always keep this section last:
## Discussion Log And User Interest
Update policy:
- Create the note on the first substantive paper question.
- Revise synthesis sections as understanding improves.
- Do not preserve outdated interpretations except by correcting them.
- Keep
Discussion Log And User Interest append-only, except typo fixes.
- Add concise log bullets for every substantive discussion turn.
- Do not add citation-alert sections or email drafts to paper notes.
- Add future ideas only when they naturally emerge or the user asks for them.
Chat Style
Answer the user's current question directly. Do not force a fixed template in chat. Use the note for continuity, not for verbose terminal output.
Distinguish clearly between:
- What the paper proves or demonstrates.
- What the paper suggests.
- Your inference or interpretation.
- Why it matters for the user's research.
Feedback Integration
At the end of each substantive paper-discussion turn, run the feedback-memory checklist.
Write feedback when:
- The user gives explicit
+1, +2, -1, -2, or evaluative language.
- The user asks several substantive follow-up questions about the same paper.
- The user connects the paper to their research interests.
- The user asks for concrete future ideas or actionability.
Scoring convention:
score: 1: Several substantive follow-ups or mild positive interest.score: 2: Explicit strong praise, high priority, or strong reusable preference.score: -1: Skeptical or weak negative judgment.score: -2: User concludes the paper is trivial, cheating, wrong, not useful, or should be avoided.
Several rounds of substantive discussion count as weak positive interest by default unless the user shows an obvious negative attitude. Purely logistical questions do not count.
Use feedback-memory for durable idea records. Paper notes may contain lightweight possible ideas, but user_profile/ideas.local.jsonl should only be created or updated through the feedback-memory workflow.
