research

par google-gemini

Rassemblez le matériel source pour l'émission de radio à partir de toute source de contenu — web, HN, GitHub, articles ou sujets généraux.

npx skills add https://github.com/google-gemini/gemini-managed-agents-templates --skill research

Télécharger ZIP GitHub

Research

Gather source material for the radio show based on the user's prompt. The research strategy is agent-driven — you decide how to gather content based on what the user asks for.

Scripts

All research scripts live in skills/research/scripts/ and output to {workspace}/data/research/. They use only Python stdlib — no dependencies.

Script	Command	Source
`fetch_hn.py`	`python3 skills/research/scripts/fetch_hn.py --workspace ./workspace --top 6`	Hacker News
`fetch_github.py`	`python3 skills/research/scripts/fetch_github.py --repo owner/repo --workspace ./workspace`	GitHub repos
`fetch_url.py`	`python3 skills/research/scripts/fetch_url.py --url https://... --workspace ./workspace`	Any URL

Script Details

fetch_hn.py — Hacker News (two-phase)

Phase 1: Scan — grab top 10 story titles, scores, and comment counts (fast, no deep-diving):

python3 skills/research/scripts/fetch_hn.py --workspace ./workspace --mode scan --top 10

Output: {workspace}/data/research/hn-scan.md — a list of stories with IDs.

You read this file, pick the 2-3 most interesting stories for radio, then run phase 2.

When picking stories, SKIP anything related to politics, race, religion, international conflicts, historical controversies, gender/culture wars, or immigration. Stick to tech, programming, AI/ML, open source, science, startups, and developer culture.

Phase 2: Deep-dive — fetch full comment threads for the stories you picked:

python3 skills/research/scripts/fetch_hn.py --workspace ./workspace --mode deep-dive --stories 43210987,43209876

Output: {workspace}/data/research/hacker-news.md — full stories with top comments.

Argument	Default	Description
`--workspace`	`workspace`	Root workspace directory
`--mode`	(required)	`scan` or `deep-dive`
`--top`	`10`	Number of stories to scan (scan mode only)
`--stories`	(required for deep-dive)	Comma-separated story IDs

fetch_github.py — GitHub Repository

python3 skills/research/scripts/fetch_github.py --repo googleapis/python-genai --workspace ./workspace

Accepts owner/repo or a full GitHub URL (https://github.com/owner/repo).

Argument	Default	Description
`--repo`	(required)	GitHub repo (owner/repo or full URL)
`--workspace`	`workspace`	Root workspace directory
`--releases`	`5`	Number of releases to fetch
`--issues`	`8`	Number of top issues to fetch

What it does:

Fetches repo metadata (stars, description, language).
Fetches and decodes the README.
Fetches recent releases with changelogs.
Fetches top issues by comment count, including top comments.
Outputs → {workspace}/data/research/github.md

Uses the GitHub REST API directly — no auth needed for public repos.

fetch_url.py — Any URL

python3 skills/research/scripts/fetch_url.py --url https://example.com/blog-post --workspace ./workspace

Argument	Default	Description
`--url`	(required)	URL to fetch content from
`--workspace`	`workspace`	Root workspace directory
`--max-chars`	`8000`	Max chars to extract

What it does:

Fetches the HTML page.
Strips scripts, styles, nav, footer.
Converts HTML to markdown-like text.
Extracts title and meta description.
Outputs → {workspace}/data/research/url_<safe_name>.md

Works for blog posts, documentation pages, arXiv abstracts, news articles, etc.

General Topic (no script — agent-driven)

If the user provides a topic without a specific source:

Use Google Search to find recent articles, blog posts, and discussions
Gather multiple perspectives and opposing viewpoints
Write the research markdown directly to {workspace}/data/research/

Output

Directory: {workspace}/data/research/
Format: One or more markdown files with structured content
All research must be saved here regardless of source — the script-writing step reads from this directory.