Flux (https://flux.yoandev.co) aggregates every day the articles, podcasts and videos from 80+ French-language and international tech sources. This skill provides access to this corpus for personalized watch, morning briefs and ad-hoc research.
Philosophy
The skill provides a discovery primitive (a bash script that queries the Flux index). Everything else — semantic filtering by theme, editorial selection, synthesis — is done by the agent in natural language. No arguments --themes to memorize, no config: the user speaks normally, the agent composes.
Flux never stores the full content of articles, only the metadata (title, snippet, URL, date, source, categories). To read the substance of an article, you have to fetch its link with a fetch tool (WebFetch, curl, equivalent).
Prerequisites
curlandjqin the PATH. They are installed by default on macOS and most Linux distributions. Otherwise:brew install jq/apt install jq.- A web page fetch tool. In order of preference:
WebFetch(native Claude Code),curl, or any equivalent available in the agent.
The discovery primitive
The script scripts/flux-discover.sh exposes three commands that always return JSON.
scripts/flux-discover.sh latest [--days N] [--lang fr|en|both] [--limit N]
scripts/flux-discover.sh search "<query>" [--days N] [--lang fr|en|both] [--limit N]
scripts/flux-discover.sh sources [--lang fr|en|both]
Options:
--days N— time window (default7)--lang—fr= French-language,en= international,both= both (defaultboth)--limit N— max number of articles (default100, set0for unlimited)
Each article returned contains:
| Field | Description |
|---|---|
title |
Article title |
description |
Short snippet (150-500 chars), not the full content |
link |
Public URL of the article — to be fetched for the full content |
pubDate |
ISO 8601 UTC |
source |
Readable name of the source (e.g.: "Simon Willison's Weblog") |
sourceUrl |
URL of the original RSS feed (rarely useful directly) |
categories |
Array of thematic tags (e.g.: ["Programmation", "IA"]) |
type |
"blog" / "podcast" / "youtube" |
lang |
"fr" or "en" |
audioUrl, duration |
Present for podcasts |
videoId |
Present for YouTube videos |
image |
URL of the preview image (optional) |
Retrieving the full content of an article
When the user wants a summary, a brief or an analysis, you have to fetch the actual content of the article (not just the description, which is only a snippet).
For type: "blog" — fetch the link with the available tool, in this order:
WebFetchif available (Claude Code) — directly gives the text/markdown extracted cleanly from the page. Recommended path.curl -sL <link>as a fallback, then extraction of the main content (either by the model's intelligence on the raw HTML, or via a readability tool if the environment has one).- If both fail (paywall, site down, anti-bot): tell the user that the content could not be retrieved and settle for the
description+link. Never invent the content of an article that could not be read.
For type: "podcast" or type: "youtube" — no need to fetch anything. The description already contains the useful summary (episode description, video description). Systematically offer the link to listen to/watch, and audioUrl for podcasts you want to embed directly.
Pattern: personalized morning brief
When the user asks "make me my brief", "my morning newsletter", "what happened this week on X", etc.:
- Call
flux-discover.sh latestwith a consistent window:--days 1to--days 7depending on the request. Reasonable default:--days 3. - Filter semantically the articles according to the themes mentioned by the user. Do not settle for a flat textual match — use
title,descriptionandcategoriesto judge the real relevance. An article "AI-generated code quality" is relevant to the theme "AI" even if the word "AI" does not appear. - Select 5 to 8 articles, favoring diversity of sources (avoid 5 articles from the same blog). Mix blogs + podcasts + videos when relevant.
- For each article
blogretained : fetch thelink(WebFetch preferred) to read the real content before synthesizing. - For the podcasts/videos retained : rely on the
description, without fetching. - Compose a brief in markdown structured by requested theme, with for each article:
- Title as a clickable link to
link - 2-4 sentences that summarize the substance (not just the beginning of the article)
source,typeand date in parentheses at the end
- Title as a clickable link to
- End with a section "Other readings" that quickly lists 3-5 relevant articles not detailed (title + link + source).
Pattern: ad-hoc search
When the user asks "find me articles on X" or "what has been said recently about Y":
flux-discover.sh search "X" --days 30(a wider window than a brief).- If few or no results, broaden to
--days 90or do a search by related keyword. - List the results: title, source, date, link. Group by
typeif useful. - Offer: "would you like me to go deeper into one or more of these articles?"
If no relevant results: say so honestly. Do not synthesize from nothing.
Pattern: deep dive
When the user points to an article (via link or reference):
- Fetch the
link(WebFetch preferred). - Write a structured synthesis: context, key points, conclusion, possible notable excerpts. Keep the clickable link to the original.
Pattern: landscape exploration
When the user asks "what are the French-language sources on DevOps?" or "who publishes on AI in French?":
flux-discover.sh sources --lang frto get the deduplicated list.- Cross-reference with
flux-discover.sh search "<thème>" --days 90to identify who has actually produced on the topic recently.
Cross-cutting principles
- Cite sources as clickable links : the user must be able to return to the original article in one click.
- Respect the output language : by default, respond in the user's language (often French). For an English article summarized for a French speaker, translate the summary.
- Honesty about failures : if a fetch fails, say so, move on to the next article. Never hallucinate the content of an inaccessible article.
- No redundancy : if two articles cover the same topic, detail only one (the most complete or the most recent), mention the other briefly.
- Latency to be aware of : Flux runs on daily indexing (~4 UTC). Articles published after this run appear the next day. Do not promise the user minute-level freshness.
Troubleshooting
- The script returns
[]: either nothing matches (broaden--days, loosen the search term), or the endpoint is unavailable (rare — Netlify). Check withcurl -I https://flux.yoandev.co/search-index.json. jq: errorparsing: an article has a malformedpubDate. The script silently ignores it, this is not blocking.- A site blocks the fetch: note the source in the summary ("Article inaccessible via direct fetch — see link above"), continue.