Imported from Jurgita-DS/data-science-research-lab-with-google-adk (
.agents/skills/search-optimization/SKILL.md). Install upstream withnpx skills add Jurgita-DS/data-science-research-lab-with-google-adk --skill search-optimization. Copyright stays with the author.
--- name: search-optimization description: Guidance for the Search Agent on free-sources-first strategy, Google Search query construction biased toward free ML/DS content sources, and lightweight Phase 1 discovery output.
Search Agent Optimization Skill
The Search Agent uses Google Search (not the arXiv API) as its primary discovery mechanism. The core principle: bias all searches toward freely accessible content.
Free Sources Priority Tier
In ML/data science, nearly every important paper has a free version. Prioritise in this order:
| Priority | Source | Why |
|---|---|---|
| 1st | arxiv.org | Free PDFs for almost all ML papers |
| 2nd | openreview.net | Free PDFs for NeurIPS, ICLR, ICML, ICLR papers |
| 3rd | paperswithcode.com | Links papers to arXiv + GitHub implementations |
| 4th | towardsdatascience.com | High-quality applied DS/ML articles |
| 5th | medium.com | Broad ML content |
| 6th | huggingface.co | ML engineering articles and model cards |
| 7th | distill.pub | Excellent visual deep learning explainers |
| 8th | github.com | Paper implementations and README explanations |
Paywalled sources to AVOID: dl.acm.org, ieeexplore.ieee.org, link.springer.com, onlinelibrary.wiley.com, www.nature.com (unless open access)
Google Search Query Construction
Standard search query format:
[user topic] site:arxiv.org OR site:paperswithcode.com OR site:openreview.net OR site:towardsdatascience.com OR site:medium.com OR site:huggingface.co
For specific paper lookup:
"[exact paper title or author + keyword]" site:arxiv.org OR site:openreview.net
For implementation guidance:
[technique] implementation tutorial site:towardsdatascience.com OR site:medium.com OR site:huggingface.co OR site:github.com
Phase 1 Discovery Rules
Phase 1 is lightweight -- NO downloads, NO full content fetching.
For EACH search result:
- Title: from the search result title
- Source: determine from the URL domain
- Year: from the search snippet or metadata (use "N/A" if not available)
- Description: EXACTLY ONE sentence from the search snippet or abstract -- do NOT follow the URL
- URL: the direct link
De-duplication rule: if a paper appears in both the user library and web search, show it ONCE as "Internal Library" (the local copy takes precedence).
Library Check First
ALWAYS call search_library_items before running google_search. If the library has highly relevant results, show those at the top with the Internal Library label.
Result Limits
Aim for 5-10 results total. Quality over quantity. Avoid including paywalled results unless no free version can be found. When including a paywalled result, always note it and suggest the arXiv search URL.