Skip to content
Back to search

paper-finder

SkillGitHubIncomplete

Claude skill for finding ML research papers.

Safety scan

A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.

Checked today
Badge for your README
findsafeskills safety rating

Markdown

[![findsafeskills safety rating](/api/listings/bchao1--paper-finder/badge.svg)](/listings/bchao1--paper-finder)

HTML

<a href="/listings/bchao1--paper-finder"><img src="/api/listings/bchao1--paper-finder/badge.svg" alt="findsafeskills safety rating" /></a>

Install

i

What was scanned

The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.

SKILL.md · Manifest · 126 lines

View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)

  1. ---
  2. name: paper-finder
  3. description: "Finds and organizes ML, CV, NLP, and AI research papers based on textual descriptions and keywords. Searches across arxiv, Google Scholar, Semantic Scholar, and top venues (CVPR, ECCV, ICCV, NeurIPS, ICML, ICLR, SIGGRAPH, SIGGRAPH Asia, EMNLP, ACL, CoRL, RSS, AAAI, IJCAI, WACV, BMVC, MICCAI, KDD, WWW, NAACL, COLM, and more). Maintains a persistent memory bank of discovered papers, a mind-graph linking papers to topics, individual paper summaries, and BibTeX entries. Use this skill whenever the user wants to find papers, search for related work, build a literature review, discover what exists on a topic, compare papers, organize references, generate BibTeX, or manage a research paper collection. Also use when the user mentions 'find me papers on...', 'what papers exist about...', 'related work for...', 'literature search', 'paper survey', or references any ML conference/venue by name."
  4. ---
  5. # Paper Finder
  6. Research paper discovery and organization agent. Find relevant ML/AI/CV/NLP papers, organize them into a persistent knowledge base, and connect them across topics.
  7. ## Directory Structure
  8. Each search/topic gets its own folder. The folder name should be a short, descriptive kebab-case name for the search topic (e.g., `mixed-resolution-diffusion/`, `video-generation-efficiency/`). The user may also specify a custom folder name. Create on first use:
  9. ```
  10. <topic-name>/
  11. memory-bank.md # Master list of all discovered papers
  12. mind-graph.md # Topic-paper connection graph
  13. summaries/ # Per-paper .md files (via research-paper-analyst skill)
  14. references.bib # Combined BibTeX for all papers
  15. pdfs/ # Downloaded PDFs (only when user asks)
  16. discussions/ # Paper comparison logs
  17. ```
  18. If the user references an existing folder (e.g., `@mixed-resolution-diffusion/`), operate within that folder. If starting a new search without a specified folder, derive a descriptive name from the search query.
  19. ## Searching for Papers
  20. ### Web search is mandatory
  21. Use WebSearch and WebFetch for every search. Training knowledge alone misses recent papers (2024-2025+). If web tools are denied, retry once, then tell the user you need web access and explain what you'd search for.
  22. ### Search strategy
  23. Run 2-3 parallel searches per query:
  24. 1. **Semantic Scholar API** via WebFetch: `https://api.semanticscholar.org/graph/v1/paper/search?query=<query>&limit=20&fields=title,authors,year,venue,abstract,externalIds,citationCount,url`
  25. 2. **WebSearch** with queries like `<topic> paper <venue> <year>` — good for Google Scholar results
  26. 3. **Venue-specific** when relevant: `<topic> CVPR 2025`, `<topic> site:openreview.net`
  27. 4. **Follow citations** on Semantic Scholar for highly relevant papers
  28. Relevant venues by field: CV (CVPR, ECCV, ICCV, WACV), ML (NeurIPS, ICML, ICLR, COLM, AAAI), NLP (ACL, EMNLP, NAACL), Graphics (SIGGRAPH, SIGGRAPH Asia, 3DV), Robotics (CoRL, RSS, ICRA), Medical (MICCAI), Preprints (arXiv cs.CV/CL/LG/AI).
  29. ### Multi-angle search (mandatory)
  30. A single concept can be described using very different vocabulary depending on the angle. After the initial direct-concept searches, you MUST run at least one additional search round covering these three angles. Skipping these is the #1 cause of missed papers.
  31. 1. **Cross-domain synonyms**: The same idea often has established names in adjacent fields. Before searching, brainstorm 2-3 alternative terms from related domains (graphics, neuroscience, signal processing, HCI, information theory, etc.). For example, "mixed-resolution spatial tokens" in ML maps to "foveated rendering" in graphics, "saliency-driven attention" in neuroscience, or "non-uniform sampling" in signal processing. Search using these alternative vocabularies.
  32. 2. **Enabling mechanisms / building blocks**: Search for the specific technical components needed to *implement* the concept — not just the concept itself. Every novel representation requires changes to attention, positional encodings, loss functions, normalization, etc. For example, mixed-resolution tokens require modified RoPE/positional embeddings, cross-resolution attention alignment, and boundary handling. Search for these mechanism-level terms (e.g., "positional encoding mixed resolution," "RoPE phase alignment multi-scale").
  33. 3. **Motivating applications / problem framing**: Papers solving the same technical problem may frame it as a different goal. Search from the perspective of *why* someone would build this (efficiency, speed, perceptual quality, hardware constraints). For example, "spatial acceleration diffusion" and "latent upsampling" lead to mixed-resolution tokens as a solution, but would never surface from searches for "mixed-resolution tokens" directly.
  34. After initial results come in, also **follow the citation graph**: fetch the related-work section of 1-2 top-relevance papers and scan for references you haven't found yet.
  35. ### Understand the concept precisely
  36. Before searching, understand the exact technical distinction the user cares about. If they describe a specific mechanism (e.g., "tokens of different spatial sizes within a single image"), search for that literal property — don't broaden to superficially similar but technically different work (e.g., cascaded pipelines, super-resolution).
  37. ### Filtering
  38. - **Prioritize algorithmic contributions** over architecture/engineering/systems papers
  39. - **Prioritize recent work** (2024-2025+) — skip well-known basics (DiT, VQGAN, etc.) unless directly relevant
  40. - **Note citation counts** when available
  41. - **Tier results** by relevance to the user's specific concept
  42. ## Memory Bank (`memory-bank.md`)
  43. Master record of all discovered papers. Append new entries, never overwrite. Read existing file before searching to avoid duplicates.
  44. ```markdown
  45. # Paper Memory Bank
  46. Last updated: YYYY-MM-DD
  47. ### [short-id] Paper Title
  48. - **Authors**: Author list
  49. - **Venue**: Conference/Journal, Year
  50. - **URL**: Link to paper
  51. - **Citations**: N (if known)
  52. - **Status**: discovered | summarized | analyzed
  53. - **Topics**: topic1, topic2
  54. - **Abstract**: 1-2 sentence description
  55. - **Notes**: Relevance observations
  56. ---
  57. ```
  58. ## Mind Graph (`mind-graph.md`)
  59. Topic-centric hierarchy — NOT pairwise paper comparisons. Each topic has 1-3 umbrella/landmark papers plus other relevant work.
  60. ```markdown
  61. # Mind Graph
  62. Last updated: YYYY-MM-DD
  63. ### Topic Name
  64. - **Description**: One-line description
  65. - **Related topics**: [other topic], [other topic]
  66. - **Key papers**:
  67. - [short-id] Paper Title (Venue Year) — why it's key for this topic
  68. - **Other relevant papers**:
  69. - [short-id] Paper Title — one-line note
  70. ```
  71. ## BibTeX (`references.bib`)
  72. Write a single combined `references.bib` file with all papers. Use `@inproceedings` for conferences, `@article` for journals, `@misc` for arXiv preprints. Citation key = short-id.
  73. ## Paper Summaries and Comparisons
  74. - **Summaries**: Invoke **research-paper-analyst** skill. Save to `summaries/<short-id>.md`. Only when user explicitly asks — don't auto-summarize.
  75. - **Comparisons**: Read existing summaries first (create if missing via research-paper-analyst), save discussion to `discussions/<descriptive-name>.md`.
  76. - **References to known papers**: Search summaries and memory bank first. Only re-read the original paper if the user explicitly asks.
  77. ## PDF Management
  78. Do NOT download PDFs unless the user explicitly asks. When asked:
  79. 1. **Read `references.bib`** to extract the arXiv eprint ID or URL for each paper. This is the canonical source — do NOT read memory-bank.md or other files just to find download URLs.
  80. 2. Construct the PDF URL from the arXiv ID: `https://arxiv.org/pdf/<eprint-id>`
  81. 3. Download via curl/WebFetch and save to `pdfs/<short-id>.pdf`
  82. 4. Only fall back to memory-bank.md or web search if a paper has no entry in references.bib.
  83. ## Interaction Flow
  84. 1. **Search**: Run parallel web searches, present ranked list (title, venue, year, citations, one-line description)
  85. 2. **Record**: Add papers to memory-bank.md, update mind-graph.md, write references.bib
  86. 3. **Ask**: Whether user wants deeper analysis of any specific papers

Ratings and comments

Sign in to rate this skill.

Sign in

Free to use. If it’s helped, pass it on.