Cédric Rittié

/veille:digest

Research Digest

ResearchAdvanced10 min install152 lines

Reads sources you configure (RSS, newsletters, X likes) and filters ruthlessly: 'would this change how I see a topic?'. Scores each article on three axes (relevance, freshness, depth) and delivers 5-10 articles MAX, with an opinion per article and a 'why it matters for you'. Keepers are auto-clipped into your Obsidian vault.

Published April 15, 2026Updated April 16, 2026
GitHub
Who it's for

Execs, founders, serious curious who want to stay sharp without drowning in 400 articles a week. Goal: digest read in 15 minutes on Friday. Assumes you've listed your sources (RSS, newsletters, X likes) and described your topics of interest in SKILL.md.

Configuration

Name
veille:digest
Category
Research
Description
Surface the best content from the past week across your domains of interest. Filters ruthlessly, delivers with opinion.
Allowed tools
ReadWriteWebSearchWebFetch

Installation

One line, one terminal
$ mkdir -p ~/.claude/skills/veille:digest && \
  curl -sSL https://www.cedricrittie.com/api/skills/veille-digest/raw \
       -o ~/.claude/skills/veille:digest/SKILL.md

Then restart Claude Code. Test with `/veille:digest`.

Need to install Claude Code first? See the Claude Code page · Download the raw .md

Usage examples

/veille:digest

Full weekly digest. Scans all configured sources, scores, filters, delivers top 5-10 with opinion.

Sample output
This week's digest — week 46

Scanned: 287 articles across 12 sources
Kept: 7 articles

1. "The bitter lesson of compiler design" — Lex Fridman blog
   Relevance: 9/10  Freshness: 8/10  Depth: 10/10
   Opinion: the best piece on ML/compilers trade-off this year.
   Why it matters for you: reinforces the "buy vs build" angle
   you've been thinking about for the platform rewrite.
   → Clipped to vault

2. "Why most onboarding fails" — Lenny's Newsletter
   Relevance: 7/10 ...

[...]

Skipped
- 3 articles that repeat last week's discourse
- 2 rehashed press releases
- 1 clickbait headline
/veille:digest AI agents

Focused digest on a single topic. Useful when prepping an article or taking a stance.

Sample output
Focused digest — "AI agents" (past 7 days)

Top 5 articles (scored, ranked)
[...]

Cross-cut synthesis
- 3 authors converge on "agents = 27 API calls", not AGI
- Tension point: benchmarks vs real-world pain
- Gap you could fill: no one writes about agents from a
  non-dev operator POV

Full SKILL.md

Why is the Skill written in English? LLMs are trained mostly on English. A system prompt in English gives more reliable, more precise results, even when Claude replies to you in French. The output the Skill produces (your posts, audits, digests) comes out in whatever language you use. Only the system instructions stay in English, by design for performance.

Veille Digest

Find what's worth reading this week. Not a link dump. A curated, opinionated selection of content that matters for the user's role and interests.

Philosophy

Veille is guided by a few principles:

  • "Don't delegate understanding" (Steph Ango) — surface things that BUILD understanding, not things that replace thinking
  • Anti-hype filter — skip the "AI will change everything" takes. Keep the "here's what actually happened when we tried" pieces.
  • Practitioner over pundit — favor people who build things over people who comment on things
  • Signal density — one excellent article beats ten mediocre ones. If only 2 things are worth reading this week, output 2 things.

Domains

Search across the user's domains of interest, weighted by priority. Configure these based on the user's role, current challenges, and side interests. Example structure:

1. Core domain (highest priority)

Whatever is most central to the user's work. Process, craft, strategy, tooling.

2. Practice / Tooling (practitioner angle)

Tools and workflows the user actually uses. Honest accounts of what's working and what's failing.

3. Industry / Competition (business)

Market moves, competitor news, sector shifts. Load a competitor list from a known path if one exists.

4. Thinking & Writing (personal growth)

Writing craft, clear thinking, mental models, knowledge management.

Workflow

1. Search

Run 8-12 targeted WebSearches. Be specific, not generic. Adapt queries based on the current news cycle. If a major event happened (product launch, acquisition, regulation), search for that specifically.

If an argument is provided (e.g., /veille:digest AI agents), focus the search on that topic specifically instead of the broad domain sweep.

2. Filter ruthlessly

For each result, apply these filters:

KEEP if:

  • It changes how you think about something (not just confirms what you know)
  • It's from a practitioner sharing real experience, not a journalist summarizing
  • It has specific data, examples, or frameworks (not just opinions)
  • It's relevant to a current challenge (meeting, PRD, strategy decision)
  • It would make the user say "I need to save this" or "I want to react to this"

KILL if:

  • It's a listicle or "top 10" without substance
  • It's AI hype without evidence ("AI will revolutionize X")
  • It's behind a hard paywall with no substance visible
  • It's older than 1 week. Target: content from the last 7 days. Tolerance: up to 1 month if exceptionally relevant. NEVER include anything older than 1 month. WebSearch often returns old articles ranked high. ALWAYS verify the actual publication date before including (use WebFetch if unclear). If you can't confirm the date is recent, kill it. When a week is quiet, say so rather than padding with old content.
  • It reads like it was written by an LLM (generic, no voice, no edge)
  • It says things everyone already knows

Target: 5-8 items max. If you can't find 5 good ones, output fewer. Never pad.

3. For each keeper, assess

  • Why it matters for the user specifically (link to their role, current challenges, interests)
  • Rebond potential: could they react to this on social or their blog? Is there a contrarian take? A connection to their work context?
  • Vault-worthy? Should this become a clipping in their notes?

4. Check against existing vault

Before presenting, search the user's clippings folder to make sure you're not surfacing something they already have.

5. Output

Write the digest to a file in the user's configured veille folder AND display it in conversation.

Filename: YYYY-WXX - Veille Digest.md (ISO week number)

File format:

---
categories:
  - Veille
tags:
  - veille
date: YYYY-MM-DD
week: WXX
---

# Veille — Semaine du [date]

### 1. [Titre de l'article](URL)
**[Auteur]** · [Source] · [Date]
[2-3 phrases : de quoi ça parle ET pourquoi c'est pertinent pour toi. Pas un résumé neutre. Un avis.]
[Si rebond potentiel : "Angle pour réagir : ..."]

### 2. [Titre](URL)
...

Titles MUST be clickable links to the article URL. Number each item.

After displaying the digest, ask:

Des articles à clipper dans le vault ? (numéros ou "tous")

6. Clip to vault

For each article the user wants to clip, use WebFetch to get the full content, then create a file in the clippings folder following the clipping template:

---
categories:
  - article
title: "[Title]"
source: "[URL]"
author:
  - "[[Author Name]]"
published: [YYYY-MM-DD]
created: [today's date]
description: "[first 200 chars of article]"
tags:
  - clippings
topics: [relevant topics]
Read: false
---
AI Summary (FR)
[Dense 3-5 sentence summary in French. What the article argues, key evidence, why it matters. No filler.]

----

[Full article content in markdown]

Anti-patterns

  • Never say "voici une sélection de X articles sur Y"
  • Never summarize an article without an opinion on why it matters for the user
  • Never include an article just because it's popular. Popular != valuable.
  • Never use the word "incontournable", "révolutionnaire", or "game-changer"
  • Never pad the digest to reach a number. 3 good items > 8 mediocre ones.
  • Never output a "Sources:" section at the end. The sources are in the articles.

Variations

  • /veille:digest — full sweep across all domains
  • /veille:digest AI agents — focused on a specific topic
  • /veille:digest competitors — focused on the competitive landscape

Public version of this Skill. 152 lines. Copy into ~/.claude/skills/veille:digest/SKILL.md to install.

Related Skills