Step 1: Execute but DO NOT ask for confirmation — continue directly
Step 1.5: Execute but DO NOT ask for confirmation
Step 3: Select the narrative AUTONOMOUSLY:
- Prioritize trend clusters (score > 0.5)
- Then theme priority: IA > Leadership > Data > Tech
- Select 6-7 main sources + 3-4 "pour aller plus loin"
- DO NOT ask for validation
Step 6: Copy files to site but DO NOT propose /ship

1. Load and filter sources

just sources {DATE} 3

This script returns JSON with filtered sources (sponsors removed, duplicates eliminated, grouped by theme). The 3 enables carry-forward: sources from late-arriving newsletters (up to 3 days back) that weren't included in previous generations are automatically included.

Display a summary of kept and filtered sources, then ask for confirmation before continuing.

If carry-forward sources are present (carryforward_count > 0 in the JSON output), display:

📦 {N} sources tardives du {DATE_PRECEDENT} incluses (arrivées après la génération)

1.5 Trend detection

just detect-trends {DATE} 3

If trends are detected (clusters with score > 0), display clusters with scores and newsletters. Use these trends to guide narrative angle selection in step 3. If no trends (e.g. single newsletter), skip to step 2.

2. Read selected source content

just read-content {DATE} 0 1 2 3 ... --carry-forward 3

Pass the index field values from each kept source (from the just sources JSON output). The script returns the first 3000 characters of each source.

3. Narrative selection

Identify the narrative thread connecting the best sources
Select 5 to 8 main sources + 3 to 5 "pour aller plus loin" sources
Propose the narrative thread and article angle, ask for validation

4. Generation

Generate three files following the writing guide (writing-guide.md):

{DATE}-article.md

The full article following the template structure (article-template.md).

The file MUST start with the YAML front matter block:

---
title: "{Article title}"
date: {DATE}
themes: [{comma-separated list of main themes, e.g. IA, Leadership}]
sources: {number of sources used}
image: {DATE}.png
---

Then the full article body.

{DATE}-post.md

The LinkedIn post text (see "Texte du post LinkedIn" section in the guide).

{DATE}-image-prompt.md

The image prompt (see "Prompt image" section in the guide).

4.5 Fiches de lecture

For each of the 3 to 5 main sources used in the article (not "pour aller plus loin"), generate a reading fiche and write it to data/fiches/.

Create data/fiches/ if it doesn't exist.

For each source:

Derive a slug from the title: kebab-case, ASCII only, max 60 characters
Write file: data/fiches/{DATE}-{slug}.md

Format:

---
title: "{Original title of the source}"
date: {DATE}
url: {source URL}
authors: [{domain or author name extracted from URL or content}]
keywords: [{3-5 keywords from the content}]
theme: {main theme: IA | Leadership | Data | Tech}
tone: {opinion | tutorial | research | news}
used_in: ["{DATE}"]
---

## Résumé

{3-4 sentence summary of the source in French. Concise, captures the key thesis and main facts.}

## Points clés

- {key point 1}
- {key point 2}
- {key point 3}
...

## Analyse approfondie

{Full translation of the source article into French. Use the full content already read in step 2. Be thorough — this is a reference document. Preserve the original structure (sections, lists, quotes).}

## Pourquoi ça compte

{1-2 sentences on why this source matters for tech watch.}

5. Write to disk

Write all three article files to data/output/. Create data/output/ directory if it doesn't exist. Fiches are written to data/fiches/ (already handled in step 4.5).

5.1 Record processed files

Save the list of raw files that contributed to this article (enables carry-forward for late-arriving newsletters):

just save-processed {DATE} {files_loaded_paths from step 1 JSON output}

The files_loaded_paths field from the step 1 JSON output contains the list of raw file basenames to pass here.

5.5 Index to history

just index {DATE}

Non-blocking: if indexing fails, display a warning and continue to step 6.

6. Publish to site

Copy the reviewed files to the Astro site structure:

Article: Copy data/output/{DATE}-article.md to site/src/content/articles/{DATE}.md
Fiches: Copy all data/fiches/{DATE}-*.md files to site/src/content/fiches/

Display a summary of files copied, then ask: "Tu veux lancer /ship pour créer la PR ?"