RSS digest

The site is a SPA: WebFetch on the homepage returns nothing useful. Go through the API.

Fetch

SKILL_DIR="${CLAUDE_SKILL_DIR}/scripts"
OUT="<session scratchpad>/rss"
"$SKILL_DIR/fetch.sh" "$OUT" 30     # days, default 30

compact.tsv columns: score, feed, date, title, url, summary (300 chars). Sorted by site score, unread only, deduplicated on title. Never Read it whole when it exceeds ~300 lines: filter with rg/awk on the feed or title first.

API (if the script needs changing): GET /api/articles?days=N&limit=0 returns {total, articles[]} with id, title, url, feed_title, published (epoch), score, matched_topics, matched_keywords, top_topic, summary, bookmarked. Shape it server-side rather than with jq: fields=a,b (projection), summary_chars=N (plain text), dedupe=title, min_score, feeds=a,b / exclude_feeds=a,b (exact feed_title), e.g. feeds=Cert%20SSI to pull every advisory regardless of score. Write endpoints (/api/mark-read, /api/bookmark, /api/config/scoring, /api/rescore, /api/snooze) are never called without an explicit ask.

The site score is not the ranking

It is keyword density over matched_topics, blind to source quality. Beginner listicles from DEV.to and Hacker Moon (~90% of volume) float to the top; CERT-FR advisories and vendor CVEs sit at 1-6. Re-rank by hand:

1. Security with possible exposure, any score: CERT-FR (Cert SSI), IT Connect / vBlog CVE posts, Ars security. Keep only products we run: GitLab, Atlassian, Linux kernel/KVM, container runtimes, Docker Hub, ingress/proxies, GCP, Kubernetes. Say whether we are exposed (SaaS = patched by vendor) and mark what was not verified.
2. Deadlines: Kubernetes or GKE EOL, deprecations, forced upgrades.
3. GKE / GCP / FinOps: the Containers & Kubernetes (Google Cloud blog) feed is the highest-signal source, read all of it.
4. Engineering deep dives: Meta Engineering, Cloudflare (data pipelines, reliability, observability; skip Workers product launches), Percona only for incident-shaped posts.
5. DEV.to / Hacker Moon: keep only items naming our stack explicitly (Thanos, ArgoCD, Terragrunt, GKE, Workload Identity, PromQL, cardinality, FinOps on GCP) and written from production, not tutorials.

Everything else (consumer tech, Apple, politics, beginner Kubernetes, "AI incident agent" series) goes in one closing line.

Profile

Platform/SRE at Equativ (adtech, RTB bidders, high QPS). GCP + GKE fleet (~24 prod clusters), ArgoCD GitOps, Terraform/Terragrunt (iac-*), Prometheus + global Thanos + Grafana Cloud, Tempo, Cloudprober, Aerospike, Dragonfly, GitLab CI, Akamai edge, some OVH bare metal (CDYN/Dynadmic). FinOps-first.

Output

In French. Sections in this order, each a table or short list with a link, the site score and one line of "why it matters here": à traiter (sécu), calendrier, GKE/FinOps, lectures de fond, then one line for the rest. Lead with the backlog size and the source split. Do not mark anything read unless asked.