Upgrade AI analysis guide to v4.0 and improve all news workflow prompts for quality

.github/workflows/news-article-generator.md

@@ -631,6 +631,29 @@ For **non-deep-inspection** article types only, if the script fails, generate ar
 
 **4. Update all metadata** — Ensure `<title>`, `<meta name="description">`, `<meta property="og:title">`, `<meta property="og:description">`, and `<h1>` all reflect the AI-generated title and description.
 
+## Step 3c: AI Content Quality Enforcement (v4.0 — MANDATORY)
+
+> 🚨 **v4.0 CRITICAL**: This is the multi-type article generator. Apply content quality enforcement for ALL article types. See `SHARED_PROMPT_PATTERNS.md` §"AI ARTICLE CONTENT GENERATION" and `ai-driven-analysis-guide.md` v4.0.
+
+**1. Read pre-computed analysis** — For the current `${REQUESTED_TYPE}`, read ALL analysis files from `analysis/daily/${ARTICLE_DATE}/${ANALYSIS_SUBFOLDER}/`. If synthesis reports "0 documents analyzed", use MCP tools to fetch data directly (see ai-driven-analysis-guide.md §Empty Analysis Fallback).
+
+**2. Scan for BANNED content patterns** — Search each generated article for these exact strings or equivalent boilerplate patterns and REPLACE them:
+- Exact string: `"The political landscape remains fluid"` → Replace with specific winners/losers
+- Exact string: `"No chamber debate data is available"` → Replace with analysis from document text or MCP debate data
+- Pattern/prefix match: any `"Touches on ... policy."` boilerplate followed by generic domain text → Replace with unique per-document analysis
+- Pattern/prefix match: any boilerplate starting with `"Analysis of "`, followed by a document count and `" documents covering"` → Replace with analytical lede
+
+**3. Enforce per-document unique "Why It Matters"** — Verify that NO two documents in the same article share identical "Why It Matters" text. If found, rewrite each with document-specific evidence.
+
+**4. Enforce minimum analytical depth** — Every article MUST contain:
+- Analytical lede naming actors and political significance
+- Per-document analysis (not flat list of links)
+- Winners & Losers with named parties and evidence (≥50 words)
+- Key Takeaways with confidence labels (3-5 bullet points)
+- Analysis references section with GitHub links
+
+**5. Run self-quality check** — Score each article against the 5-dimension rubric from SHARED_PROMPT_PATTERNS.md §"Article Quality Self-Check". If any article scores below 7.0 composite, revise before committing.
+
 ## Step 4: Translate & Validate
 
 Check for untranslated Swedish content in non-Swedish articles:

.github/workflows/news-committee-reports.md

@@ -681,6 +681,30 @@ npx tsx scripts/fix-article-navigation.ts
 
 **4. Update all metadata** — Ensure `<title>`, `<meta name="description">`, `<meta property="og:title">`, `<meta property="og:description">`, and `<h1>` all reflect the AI-generated title and description.
 
+### Step 3c: AI Content Quality Enforcement (v4.0 — MANDATORY)
+
+> 🚨 **v4.0 CRITICAL**: The AI MUST read pre-computed analysis files and rewrite ALL script-generated stub content. See `SHARED_PROMPT_PATTERNS.md` §"AI ARTICLE CONTENT GENERATION" and `ai-driven-analysis-guide.md` v4.0 §"AI Article Content Generation Protocol".
+
+**1. Read pre-computed analysis** — Before modifying article content, read:
+```bash
+cat "analysis/daily/${ARTICLE_DATE}/committeeReports/synthesis-summary.md"
+cat "analysis/daily/${ARTICLE_DATE}/committeeReports/swot-analysis.md"
+cat "analysis/daily/${ARTICLE_DATE}/committeeReports/risk-assessment.md"
+cat "analysis/daily/${ARTICLE_DATE}/committeeReports/stakeholder-perspectives.md"
+```
+
+**2. Replace script-generated lede** — Find and replace any `<p class="lede">Analysis of N documents covering...` with an AI-generated analytical lede naming the most significant committee report, key actors, and political significance.
+
+**3. Replace boilerplate "Why It Matters"** — For EACH committee report entry, replace any `"Touches on {X} policy..."` boilerplate with document-specific analysis citing the committee code, policy measure, budget impact, and party positions.
+
+**4. Replace generic "Winners & Losers"** — Find and replace `"The political landscape remains fluid..."` with specific winners/losers naming parties (M, S, SD, V, MP, C, L, KD) with evidence from vote records or committee decisions.
+
+**5. Replace excuse-as-analysis** — Find and replace `"No chamber debate data is available..."` with either: (a) actual debate data from MCP `search_anforanden`, or (b) analysis of the committee report text itself.
+
+**6. Add Key Takeaways** — If missing, add 3-5 bullet points with bold lead phrases, dok_id citations, and [HIGH/MEDIUM/LOW] confidence labels.
+
+**7. Verify policy domain labels** — Ensure each committee report is classified by its committee code (FöU=Defence, JuU=Justice, SoU=Healthcare, etc.), NOT by keyword heuristics.
+
 ### Step 4: Translate Swedish Content & Verify Analysis Quality
 All Swedish API data MUST be translated. Check every article for `data-translate="true"` markers.
 

.github/workflows/news-interpellations.md

@@ -692,6 +692,26 @@ For each interpellation found, cross-reference the minister's response to identi
 
 **4. Update all metadata** — Ensure `<title>`, `<meta name="description">`, `<meta property="og:title">`, `<meta property="og:description">`, and `<h1>` all reflect the AI-generated title and description.
 
+### Step 3d: AI Content Quality Enforcement (v4.0 — MANDATORY)
+
+> 🚨 **v4.0 CRITICAL**: The AI MUST read pre-computed analysis and rewrite ALL script-generated stub content. See `SHARED_PROMPT_PATTERNS.md` §"AI ARTICLE CONTENT GENERATION" and `ai-driven-analysis-guide.md` v4.0.
+>
+> **Note:** This is Step 3**d** (not 3c) because interpellations has an additional Step 3b (Cross-Reference Minister Responses) and Step 3c (AI Title/Meta), shifting this enforcement step to 3d. All other workflows use Step 3c for this same enforcement.
+
+**1. Read pre-computed analysis** — Read synthesis, SWOT, risk analysis from `analysis/daily/${ARTICLE_DATE}/interpellations/`.
+
+**2. Replace script-generated lede** — Replace any `"Analysis of N documents..."` with AI lede naming the most targeted minister, the filing party strategy, and the most significant interpellation topic.
+
+**3. Replace boilerplate "Why It Matters"** — For EACH interpellation, write unique analysis citing the interpellation number, the specific question asked, the targeted minister's portfolio, and why this matters politically. BANNED: `"Touches on {X} policy..."` boilerplate.
+
+**4. Replace generic "Winners & Losers"** — Replace `"The political landscape remains fluid..."` with specific accountability analysis: which ministers face the most pressure, which opposition parties demonstrate coordination, and minister response timeliness.
+
+**5. Integrate minister response data** — Use cross-reference results from Step 3b (minister response speeches via MCP `search_anforanden`) to enrich the article with response summaries, accountability gaps, and policy commitments.
+
+**6. Replace excuse-as-analysis** — Replace `"No chamber debate data..."` with analysis from the interpellation text itself or minister response speeches.
+
+**7. Add interpellation coordination analysis** — Identify patterns: Are multiple interpellations targeting the same minister? The same policy area? Filed on the same day (suggesting coordination)?
+
 ### Step 4: Translate, Validate & Verify Analysis Quality
 
 Run validation and HTMLHint before creating PR:

.github/workflows/news-motions.md

@@ -673,6 +673,24 @@ npx tsx scripts/fix-article-navigation.ts
 
 **4. Update all metadata** — Ensure `<title>`, `<meta name="description">`, `<meta property="og:title">`, `<meta property="og:description">`, and `<h1>` all reflect the AI-generated title and description.
 
+### Step 3c: AI Content Quality Enforcement (v4.0 — MANDATORY)
+
+> 🚨 **v4.0 CRITICAL**: The AI MUST read pre-computed analysis and rewrite ALL script-generated stub content. See `SHARED_PROMPT_PATTERNS.md` §"AI ARTICLE CONTENT GENERATION" and `ai-driven-analysis-guide.md` v4.0.
+
+**1. Read pre-computed analysis** — Read synthesis, SWOT, risk analysis from `analysis/daily/${ARTICLE_DATE}/motions/`.
+
+**2. Replace script-generated lede** — Replace any `"Analysis of N documents..."` with AI lede naming the most significant opposition motion(s), filing party, and policy target.
+
+**3. Replace boilerplate "Why It Matters"** — For EACH motion, write unique analysis citing motion number, specific policy proposal, party strategy, and target government policy. BANNED: `"Touches on {X} policy..."` boilerplate.
+
+**4. Replace generic "Winners & Losers"** — Replace `"The political landscape remains fluid..."` with specific opposition strategy analysis: which parties filed, what government policies they target, likelihood of committee support.
+
+**5. Verify document count consistency** — Ensure the number of motions in the title/lede matches the number detailed in the body. If title says "50 motions" but body details 10, either expand body or correct the title.
+
+**6. Verify policy domain classification** — Each motion MUST be classified by its Riksdag committee assignment (utskott), NOT by keyword matching. BANNED: Classifying a food safety motion as "housing policy" based on co-location with housing motions.
+
+**7. Add opposition strategy context** — Explain whether motions represent coordinated opposition campaign, individual MP initiative, or response to government proposition.
+
 ### Step 4: Translate, Validate & Verify Analysis Quality
 
 Run validation and HTMLHint before creating PR:

.github/workflows/news-propositions.md

@@ -661,6 +661,24 @@ npx tsx scripts/fix-article-navigation.ts
 
 **4. Update all metadata** — Ensure `<title>`, `<meta name="description">`, `<meta property="og:title">`, `<meta property="og:description">`, and `<h1>` all reflect the AI-generated title and description.
 
+### Step 3c: AI Content Quality Enforcement (v4.0 — MANDATORY)
+
+> 🚨 **v4.0 CRITICAL**: The AI MUST read pre-computed analysis and rewrite ALL script-generated stub content. See `SHARED_PROMPT_PATTERNS.md` §"AI ARTICLE CONTENT GENERATION" and `ai-driven-analysis-guide.md` v4.0.
+
+**1. Read pre-computed analysis** — Read synthesis, SWOT, risk, and stakeholder analysis from `analysis/daily/${ARTICLE_DATE}/propositions/`.
+
+**2. Replace script-generated lede** — Replace any `"Analysis of N documents..."` placeholder with AI lede naming specific propositions, ministers, and political significance.
+
+**3. Replace boilerplate "Why It Matters"** — For EACH proposition, write unique analysis citing the proposition number (e.g., Prop. 2025/26:235), specific policy changes, budget impact (SEK amounts), and affected populations. BANNED: `"Touches on {X} policy..."` boilerplate.
+
+**4. Replace generic "Winners & Losers"** — Replace `"The political landscape remains fluid..."` with specific analysis naming government ministers who tabled the propositions and opposition parties likely to challenge them.
+
+**5. Replace excuse-as-analysis** — Replace `"No chamber debate data..."` with analysis of the proposition text itself or debate data from MCP `search_anforanden`.
+
+**6. Handle empty analysis** — If synthesis reports "0 documents analyzed", use MCP `get_propositioner(rm="2025/26")` directly. NEVER publish with "0 documents analyzed" as content.
+
+**7. Add Strategic Context** — Explain whether propositions represent coordinated government offensive (pre-election legislative push) or routine business. Cross-reference with committee reports and motions from the same date.
+
 ### Step 4: Translate, Validate & Verify Analysis Quality
 
 Run validation and HTMLHint before creating PR:

.github/workflows/news-realtime-monitor.md

@@ -815,6 +815,22 @@ If the script genuinely fails after verifying MCP, generate articles manually ON
 
 **4. Update all metadata** — Ensure `<title>`, `<meta name="description">`, `<meta property="og:title">`, `<meta property="og:description">`, and `<h1>` all reflect the AI-generated title and description.
 
+## Step 3c: AI Content Quality Enforcement (v4.0 — MANDATORY)
+
+> 🚨 **v4.0 CRITICAL**: Breaking news articles MUST have the highest content quality. Read pre-computed analysis and rewrite ALL stub content. See `SHARED_PROMPT_PATTERNS.md` §"AI ARTICLE CONTENT GENERATION" and `ai-driven-analysis-guide.md` v4.0.
+
+**1. Read pre-computed analysis** — Read ALL analysis files from `analysis/daily/${ARTICLE_DATE}/realtime-${HHMM}/` including per-document analyses in `documents/`.
+
+**2. Write intelligence-grade lede** — Breaking news ledes MUST name the specific development, key actor, quantified impact (SEK amounts, seat counts, affected populations), and urgency.
+
+**3. Write unique "Why It Matters"** per document — Each document's analysis MUST be specific to that document's content. BANNED: any repeated `"Touches on {X} policy..."` boilerplate.
+
+**4. Write substantive "Winners & Losers"** — Name specific parties, ministers, agencies, and sectors with evidence from the analysis. BANNED: `"The political landscape remains fluid..."`.
+
+**5. Include Key Takeaways** — 3-5 bullet points with confidence labels and dok_id citations.
+
+**6. Include visualization data** — For breaking news with voting data, include Chart.js vote distribution data. For defense/budget articles, include budget allocation data.
+
 ## Step 4: Validate & Translate
 
 ```bash

.github/workflows/news-week-ahead.md

@@ -491,6 +491,24 @@ npx tsx scripts/fix-article-navigation.ts
 
 **4. Update all metadata** — `<title>`, `<meta name="description">`, `<meta property="og:title">`, `<meta property="og:description">`, and `<h1>`.
 
+### Step 3c: AI Content Quality Enforcement (v4.0 — MANDATORY)
+
+> 🚨 **v4.0 CRITICAL**: Week-ahead articles require forward-looking intelligence. Read pre-computed analysis and generate prospective content. See `SHARED_PROMPT_PATTERNS.md` §"AI ARTICLE CONTENT GENERATION" and `ai-driven-analysis-guide.md` v4.0.
+
+**1. Read pre-computed analysis** — Read analysis from `analysis/daily/${ARTICLE_DATE}/week-ahead/`. If synthesis reports "0 documents analyzed", use MCP `get_calendar_events` and `get_betankanden` to populate content directly.
+
+**2. Generate forward-looking lede** — Week-ahead ledes MUST name specific upcoming events (committee votes, plenary debates, government announcements) with dates and significance. BANNED: empty or generic ledes.
+
+**3. Generate committee schedule analysis** — For each scheduled committee report debate, explain: what the committee decided, which parties filed reservations, and what the expected plenary vote outcome is.
+
+**4. Generate government agenda preview** — List upcoming government actions (propositions expected, ministerial meetings, EU engagements) with political significance context.
+
+**5. Replace generic filler** — Remove `"The political landscape remains fluid..."` and replace with specific forward indicators derived from MCP data (e.g., `get_calendar_events`, `get_betankanden`). Each indicator MUST name a real upcoming event, committee, or deadline extracted from the data — e.g., "Watch: `<COMMITTEE>` scheduling `<TOPIC>` follow-up by `<DATE from calendar>`". Do NOT hard-code example dates or event names; always source them from the current week's MCP query results.
+
+**6. Verify document count consistency** — Ensure report counts are consistent across title, lede, body, and key takeaways. Contradictory counts (17 vs 42 vs 16) are REJECTED.
+
+**7. Handle Easter/recess periods** — When parliament is in recess, explain what legislation is pending for the return session and what government agencies are acting during recess.
+
 ### Step 4: Translate, Validate & Verify Analysis Quality
 
 Run validation and HTMLHint before creating PR: