Ralph Content

Core Principle: Multi-Model Orchestration

Claude (orchestrator) makes ALL decisions:

• What research queries to run
• Whether research is sufficient
• When content quality passes
• Which diagrams need regeneration
• When to iterate vs proceed
• What prompts to send to GPT-5.4

GPT-5.4 (copywriter) generates ALL prose:

• Newsletter body copy, findings, quick hits
• LinkedIn posts, Instagram scripts
• Headlines, hooks, practitioner callouts
• Any text that will be read by the audience

Vercel AI Gateway: https://ai-gateway.vercel.sh/v1/chat/completions with model openai/gpt-5.4. Auth via AI_GATEWAY_API_KEY from .env.local.

There is NO:

• Regex parsing of outputs
• Hard-coded iteration limits
• Deterministic state machines
• Script-based orchestration

Claude reads the context files, understands the quality criteria, constructs prompts, evaluates GPT-5.4's output, and makes judgments about iteration.

Workflow

Phase 0: Load Context

Before starting, read:

1. .ralph-content/progress.txt - Prior learnings (if exists)
2. The context files in this skill's context/ directory (already loaded via skill)

Note any patterns or gotchas from previous runs.

Phase 1: Research (Perplexity Deep Research)

Goal: Gather comprehensive, verifiable information on the topic using Perplexity's deep research model.

IMPORTANT: Use Perplexity via OpenRouter for deep research, NOT the basic WebSearch tool.

How to call Perplexity:

API_KEY=$(grep OPENROUTER_API_KEY .env | cut -d'=' -f2)
curl -s https://openrouter.ai/api/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $API_KEY" \
  -d '{
    "model": "perplexity/sonar-deep-research",
    "messages": [{"role": "user", "content": "YOUR RESEARCH QUERY HERE"}]
  }' | jq -r '.choices[0].message.content'

Actions:

1. Formulate a comprehensive research query that covers:
   • Molecular/mechanistic details
   • Clinical data with study citations (author, journal, year, n, findings)
   • Latest research (specify 2024-2025)
   • Practical implications
2. Call Perplexity sonar-deep-research via OpenRouter using the Bash tool
3. Evaluate: Is the research sufficient for the content type?
   • For newsletter: Need specific studies, company examples, expert quotes
   • For lead magnet: Need mechanism details, clinical data, practical applications
4. If insufficient, run additional targeted Perplexity queries
5. Track sources for later citation

Why Perplexity Deep Research:

• Returns comprehensive, well-sourced academic content
• Includes specific study citations (author, journal, year)
• Provides mechanistic depth not available from basic web search
• Single query can return 10,000+ words of research synthesis

Quality Check (you decide):

• Do I have specific names, numbers, and timeframes?
• Do I have 3+ examples for evidence cascades?
• Do I have mechanism details, not just surface claims?
• Are my sources from peer-reviewed journals with proper citations?

Phase 1.5: Knowledge Base Enrichment (Factor Shift) — --kb flag

Goal: Surface non-obvious, cross-domain mechanistic connections from the NGM Signaling Knowledge Base that add genuine editorial value — NOT generic clinical advice.

When to use: Enabled by --kb flag. Auto-enabled for podcast_roundup. Use for ANY content that touches biological mechanisms, interventions, biomarkers, or clinical protocols. The KB contains 862 curated documents across pathways (152), interventions (413), and biomarkers (297).

Skip when: Content is purely editorial/opinion, business strategy, or has no clinical/scientific substrate.

How to call the Factor Shift Enrichment pipeline:

VS_API_KEY=$(grep VECTORSHIFT_API_KEY .env 2>/dev/null | cut -d'=' -f2 || echo "sk_fkyJyb86LyHQR9IaomQq52xCdL7a31uyYbuHDxuqoTVqdf6n")

# Write full content to temp file, use jq to safely construct JSON
cat > /tmp/kb_content.txt << 'EOF'
YOUR FULL SECTION TEXT HERE — use complete paragraphs, not summaries
EOF

jq -n --rawfile content /tmp/kb_content.txt --arg focus "YOUR ENRICHMENT FOCUS" \
  '{inputs: {content: $content, enrichment_focus: $focus}}' | \
curl -s -X POST 'https://api.vectorshift.ai/v1/pipeline/69ab22d39813d41fbf525d0c/run' \
  -H "Authorization: Bearer $VS_API_KEY" \
  -H "Content-Type: application/json" \
  -d @-

The pipeline returns (nested under outputs key):

• enriched_context — narrative markdown with sections: Mechanistic Foundation, Key Interventions & Protocols, Biomarker Landscape, Cross-Domain Connections, Evidence Gaps, Key Citations
• gap_analysis — entity extraction + gap identification

CRITICAL: Input Quality Determines Output Quality

DO NOT send compressed summaries or one-line descriptions to the KB. The pipeline performs semantic retrieval — richer input text retrieves richer KB intersections.

BAD (produces generic output):

"Exercise induces GPLD1 which benefits cognition in aged mice"

GOOD (produces specific cross-domain connections):

"Exercise is not merely calorie expenditure or mitochondrial stress. It is a secretory event
distributed across liver, muscle, adipose tissue, endothelium, and brain. GPLD1 emerged as a
liver-derived mediator capable of recapitulating some exercise-associated gains in hippocampal
neurogenesis and cognition. The canonical muscle-brain pathway remains essential: exercise
elevates lactate, which signals through SIRT1 and PGC-1α, increasing FNDC5 expression and
cleavage to irisin, which supports BDNF-linked neuroplasticity..."

Always send the FULL prose text of each section/finding/topic — multiple paragraphs with mechanistic detail. This is what enables the KB to find deep intersections rather than surface-level matches.

Enrichment Focus by Content Type

Use a rich, specific enrichment focus — not a vague category:

• Newsletter: "mechanistic depth, cross-domain connections, specific pathway crosstalk with other longevity pathways (AMPK, mTOR, NAD+, senescence), intervention protocols, biomarker interpretation, unexpected connections to other domains in the KB"
• Lead magnet: "intervention protocols and dosing, cross-domain pathway connections, biomarker panels for monitoring, evidence gaps"
• Podcast roundup: "mechanistic depth, cross-domain connections, specific pathway crosstalk, intervention protocols, biomarker interpretation, unexpected connections"
• LinkedIn: "one specific cross-domain connection or mechanistic insight that would surprise a knowledgeable clinician"
• Regulatory brief: skip (use Phase 3 verification instead)

How to Use KB Output

The key section is "Cross-Domain Connections." This is where the editorial value lives. The KB surfaces connections like:

• NAD+ constraining the exercise→SIRT1→BDNF cascade (exercise may produce less cognitive benefit as NAD+ declines with age)
• Senescent cells upregulating DPP-4, which degrades GLP-1, creating a vicious cycle between aging and sugar craving
• Five independent inputs converging on AKT-GSK-3β (menopause + insulin resistance + inactivity = catastrophic GSK-3β disinhibition)
• Metformin's AMPK activation antagonizing mTORC1, creating a real drug-exercise timing conflict
• Circadian NAMPT oscillation → NAD+ peaks → mitochondrial biogenesis windows that circadian disruption collapses

These are the insights that belong in NGM Deep Analysis callouts — specific, non-obvious, pathway-named connections that practitioners wouldn't know from the source material alone.

Generating NGM Deep Analysis Callouts

After collecting KB cross-domain connections, send them to GPT-5.4 with explicit instructions:

Write "NGM Deep Analysis" callout text. Rules:
- 2-3 sentences max
- Surface a specific cross-domain connection, not a truism
- Name specific pathways, molecules, or biomarkers
- Frame as "what practitioners wouldn't know from the source alone"
- No "For practitioners:" prefix — just state the insight directly
- Be intellectually honest about evidence confidence

DO NOT flatten KB insights into generic clinical advice like "treat exercise as a multisystem secretome." That wastes the KB's depth. Every callout should make a reader think "I didn't know that."

For podcast_roundup and multi-finding content:

Send each finding through enrichment separately with full text. Run queries in parallel. The KB returns different cross-domain connections for each topic because it retrieves against different pathway/intervention/biomarker clusters.

For LinkedIn posts with --kb:

A single KB query is usually sufficient. Extract the single most surprising cross-domain connection and use it as the post's intellectual anchor — the thing that makes someone stop scrolling. Example: "Your patient on metformin may be blunting their own exercise gains — AMPK activation from metformin antagonizes the mTORC1 signal their muscles need for protein synthesis after training."

Phase 2: Draft via GPT-5.4 + Self-Critique Loop

Goal: Generate content that passes the quality rubric using GPT-5.4 for prose generation.

Copywriting Model: All prose/copy generation is done by GPT-5.4 via the Vercel AI Gateway. Claude handles orchestration, prompt construction, research, enrichment, quality evaluation, and HTML assembly — but the actual copywriting is delegated to GPT-5.4.

How to call GPT-5.4:

AI_GATEWAY_KEY=$(grep AI_GATEWAY_API_KEY .env.local | cut -d'=' -f2)
curl -s -X POST "https://ai-gateway.vercel.sh/v1/chat/completions" \
  -H "Authorization: Bearer $AI_GATEWAY_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openai/gpt-5.4",
    "messages": [{"role": "user", "content": "YOUR PROMPT HERE"}],
    "stream": false,
    "max_tokens": 8000
  }' | jq -r '.choices[0].message.content'

Tip: For long prompts, write the prompt to a temp file first, then use jq to construct the JSON safely:

cat > /tmp/prompt.txt << 'EOF'
Your prompt here...
EOF
PROMPT=$(cat /tmp/prompt.txt) && jq -n --arg model "openai/gpt-5.4" --arg content "$PROMPT" \
  '{"model": $model, "messages": [{"role": "user", "content": $content}], "stream": false, "max_tokens": 8000}' | \
curl -s -X POST "https://ai-gateway.vercel.sh/v1/chat/completions" \
  -H "Authorization: Bearer $AI_GATEWAY_KEY" \
  -H "Content-Type: application/json" \
  -d @- | jq -r '.choices[0].message.content'

Actions:

1. Construct a detailed prompt for GPT-5.4 that includes:
   • Voice and style guidelines from context/every-voice-patterns.md
   • Content type format requirements
   • All research gathered in Phase 1
   • KB enrichment context from Phase 1.5 (if applicable)
   • Specific instructions for output format (plain text sections, no HTML)
2. Call GPT-5.4 via the Vercel AI Gateway
3. Review the output and self-critique against the rubric:
   • For newsletter: 8-point Every.to rubric
   • For lead magnet: 10-point lead magnet rubric
   • For LinkedIn: 14-point LinkedIn rubric
   • For Instagram: 10-point script rubric
   • For podcast roundup: 10-point podcast roundup rubric
4. For each failing criterion, identify what failed and where
5. If revisions are needed, call GPT-5.4 again with specific revision instructions
6. Re-evaluate. Repeat until the content passes.

Division of labor:

• GPT-5.4 writes: prose, headlines, hooks, findings, quick hits, recommendations
• Claude handles: research, KB enrichment, prompt construction, quality evaluation, HTML/JSON assembly, file management, git commits

Key Insight: The prompt you send to GPT-5.4 is everything. Include all context, all constraints, all examples. GPT-5.4 produces better copy when given a comprehensive brief — don't send vague instructions.

Phase 3: Fact Verification (if --verify flag or regulatory_brief)

Goal: Ensure claims are accurate, citations are correct, and sources are reputable.

For regulatory_brief type, this phase is MANDATORY and expanded.

Actions:

3A. Source Credibility Audit

Before verifying claims, audit ALL cited sources for credibility:

ACCEPTABLE sources:

• Peer-reviewed journals (PubMed, PMC, MDPI, Nature, Science, etc.)
• Official regulatory bodies (FDA.gov, EMA.europa.eu, WHO)
• Academic institutions (.edu domains)
• Professional medical associations
• StatPearls/NCBI Bookshelf
• Wikipedia (for general/regulatory reference only, note as such)

REJECT and replace:

• Ecommerce sites selling the product being discussed
• Supplement/peptide vendor sites (e.g., peptidesciences.com, cosmicnootropic.com)
• Biased commercial sources
• Anonymous blogs without citations
• Social media posts

If a rejected source is found, search for a peer-reviewed alternative that supports the same claim.

3B. Citation Accuracy Validation

For each citation, verify:

1. Author attribution - Correct authors listed (not misattributed)
2. Year - Publication year is accurate
3. Journal/Source - Correct journal name
4. Claim mapping - Citation actually supports the claim it's attached to

Common errors to catch:

• Wrong author (study may have multiple related papers)
• Wrong year (confusing similar studies)
• Citation supports a different claim than stated

3C. Claim Verification

1. Identify the 3-5 most significant claims in the content
2. For each claim, use WebSearch to cross-verify with primary sources
3. If a claim is inaccurate:
   • Correct it with accurate information
   • Update the source citation
4. If a claim cannot be verified:
   • Either remove it or mark it as "preliminary" with appropriate hedging

3D. Reasoning Consistency Check (for multi-document sets)

When creating related documents (e.g., multiple safety briefs), ensure:

1. Consistent evidence weighting - Same types of evidence receive similar weight across documents
2. Explicit reasoning - If conclusions differ, the justification is explicit
3. Proportional conclusions - Stronger evidence → stronger conclusions

Example: If Document A has human trial data and Document B only has preclinical data, Document A should not receive a weaker conclusion unless there's explicit justification (e.g., Document B has class-level regulatory precedent).

Phase 4: Diagram Generation + Validation

Goal: Create SVG diagrams that render correctly and illuminate concepts.

Actions:

1. Identify concepts that would benefit from visualization
2. For each diagram: a. Choose the appropriate diagram type (mechanism, comparison, process, etc.) b. Generate SVG code following context/diagram-guidelines.md c. Self-validate the SVG:
   • Parse the SVG mentally—check text positions vs container bounds
   • Verify viewBox has sufficient size for content
   • Confirm text has 40px+ padding from edges
   • Check that text inside containers is FULLY contained
   • Verify color palette compliance d. If validation fails, regenerate with specific fixes

Validation Example: "I see text at y=380 in a viewBox of height 400. With 40px padding requirement, the lowest y for text should be 360. This will overflow. I need to either increase the viewBox height or move the text up."

Phase 5: Assembly

Goal: Compose final HTML output.

Actions:

1. For newsletter: Compose email-ready HTML with embedded diagrams
2. For lead magnet: Compose full HTML document with all sections
3. For LinkedIn: Format as plain text with appropriate line breaks
4. For Instagram: Format as script with HOOK / BODY / CTA markers

Use templates from context/ngm-style-guide.md.

Phase 6: Publish + Learn

Goal: Save output and capture learnings.

Actions:

1. Save the content:

   • Newsletter: content/social-content/newsletters/YYYY-MM-DD-{slug}.html
   • Lead magnet: content/learn-platform/lead-magnets/{slug}.html + .json
   • LinkedIn: content/social-content/linkedin-posts/YYYY-MM-DD-{slug}.md
   • Instagram: content/social-content/instagram-scripts/YYYY-MM-DD-{slug}.md

2. Update .ralph-content/progress.txt:

   • What worked well
   • What needed iteration
   • Patterns discovered
   • Gotchas to avoid

3. Git commit with quality summary:

   feat: Add {type}: {title}
   
   - Quality: Passed {N}/{total} criteria on iteration {M}
   - Diagrams: {N} generated, {N} validated
   - Research: {N} sources cited
   
   Co-authored-by: factory-droid[bot] <138933559+factory-droid[bot]@users.noreply.github.com>
   

Content Type Specifics

Newsletter

• Format: Email-ready HTML
• Length: 1200-2000 words
• Diagrams: 2-3 SVG
• Quality rubric: Every.to 8-point
• Voice: Select based on content (Dan Shipper, Tina He, etc.)

Lead Magnet

• Format: Full HTML document
• Length: ~1000 words
• Diagrams: 3-7 SVG
• Additional: Mechanism table, references section
• Voice: Andrew Huberman style

LinkedIn

• Format: Plain text with line breaks
• Length: 200-300 words
• Diagrams: 1 SVG (for image post)
• Quality rubric: 14-point LinkedIn rubric

Instagram

• Format: Script with sections
• Duration: 30-90 seconds spoken
• Diagrams: None (video format)
• Quality rubric: 10-point script rubric

Podcast Roundup

• Format: Email-ready HTML (NGM editorial design)
• Length: 1500-2500 words
• Diagrams: 0 (text-driven format)
• Quality rubric: Podcast roundup rubric (see quality-rubrics.md)
• Voice: Every.to editorial — confident, specific, forward-looking
• KB Enrichment: REQUIRED (this is the core differentiator)

Podcast roundup workflow (replaces generic Phase 1-2 flow):

1. Phase 1: Episode Research — Use Perplexity to find recent episodes from the podcast roster (Huberman, Attia, Patrick, Hyman, Nicola, Campbell, and emerging voices). Extract the 5-8 most interesting, novel, or clinically actionable findings across all episodes. Do NOT force findings into predefined themes.

2. Phase 1.5: KB Enrichment (REQUIRED) — For each major finding, call the Factor Shift Enrichment pipeline with the finding as content and "mechanistic depth, pathway connections, intervention protocols, evidence quality" as enrichment focus. The KB enrichment serves three purposes:

   • Deepen: Add mechanistic specificity the podcaster may have glossed over (pathway names, molecular targets, gene symbols)
   • Connect: Surface cross-domain connections between findings that the podcasters didn't make (e.g., two episodes discussing different aspects of the same signaling pathway)
   • Challenge: Where the KB documents evidence gaps, dose-response paradoxes, or conflicting data, use those to add intellectual honesty

3. Phase 2: Draft — Write the roundup as a series of short, punchy findings (not mini-essays). Each finding:

   • 2-3 paragraphs max (not 4-5 dense paragraphs)
   • Opens with the most surprising or actionable claim from the episode
   • Weaves in 1-2 KB enrichment insights that deepen or connect the finding
   • Ends with a practitioner-relevant implication
   • Source attribution: podcaster, show, episode, date

Structure:

HEADER: Title, subtitle, date range, byline
THE BOTTOM LINE: 2-3 sentence executive summary
FINDINGS (5-8): Each is a self-contained insight enriched by KB context
  - Finding headline (specific, not thematic)
  - Episode attribution
  - The claim + KB-enriched analysis (2-3 paragraphs)
  - NGM DEEP ANALYSIS callout (what the knowledge base adds)
QUICK HITS: 3-5 one-liner mentions of other notable moments
WHAT TO LISTEN TO: Top 3 episode recommendations with reasons

What NOT to do:

• Do NOT group findings into forced themes (e.g., "The Exercise Prescription Is Being Rewritten")
• Do NOT write 4-5 paragraph mini-essays per finding — keep it tight
• Do NOT summarize episodes — extract the single most interesting claim and enrich it
• Do NOT use generic theme headlines — every headline should be specific and surprising

Output location: content/social-content/newsletters/YYYY-MM-DD-podcast-roundup.html (+ .json)

Regulatory Brief (NEW)

• Format: Print-optimized HTML (8.5" × 11", 0.75" margins)
• Length: One page (~600-800 words)
• Diagrams: 0-1 (tables preferred)
• Quality rubric: 12-point regulatory brief rubric
• Verification: MANDATORY (--verify auto-enabled)
• Voice: Formal, objective, evidence-based

Structure for regulatory briefs:

1. Header - Document type, subject, key identifiers (MW, sequence, CAS#)
2. Section 1: Safety Data Summary
   • Clinical trials (with n, design, key findings)
   • Observational/real-world data
   • Route-specific considerations (table format)
   • Adverse event profile (common/uncommon/rare)
   • Safety signals (Known/Theoretical/Absent)
3. Section 2: Regulatory Precedent
   • EMA/non-U.S. approvals
   • Historical FDA position
   • Analogous drug precedents
4. Section 3: Conclusion
   • Evidence synthesis
   • Explicit statement of evidence gaps
   • Recommendation with conditions
5. References - All with clickable links, peer-reviewed sources only

Output location: content/social-content/newsletters/YYYY-MM-DD-{slug}.html (+ .json)

Example Run

User: /ralph-content "How cold exposure affects brown adipose tissue" --type newsletter --verify

You (thinking through each phase):

PHASE 0: CONTEXT
- Reading progress.txt... Previous runs show diagram text overflow was common issue.
  I'll be extra careful with text positioning in diagrams.

PHASE 1: RESEARCH
- Query 1: "cold exposure brown adipose tissue activation mechanisms 2025 2026 studies"
  [WebSearch results with studies]
- Query 2: "BAT thermogenesis UCP1 clinical applications"
  [WebSearch results with mechanism details]
- Evaluating: I have Hanssen et al. 2015, Cypess et al. 2009, specific activation 
  temperatures, and metabolic rate increases. Sufficient for newsletter.

PHASE 2: DRAFT + CRITIQUE
- Writing draft with concrete scenario hook...
- Self-critique against 8 points:
  ✓ Specificity: Named researchers, 17.5°C threshold, 5-15% metabolic increase
  ✗ Evidence cascade: Only 2 studies stacked, need 3+
  ✓ Colon technique: Used 4 times
  ✓ Em-dash clarity: Used for definitions and contrast
  ✓ One-sentence paragraphs: 4 strategic placements
  ✓ Concrete opening: Dr. Cypess case at Joslin
  ✓ Voice: Confident, specific
  ✓ Forward-looking: Ends with implications for metabolic practice
- Revising: Adding Ouellet et al. 2012 to evidence cascade...
- Re-evaluating: All 8 criteria now pass.

PHASE 3: VERIFICATION
- Claim: "Cold exposure at 17-19°C activates BAT"
  [WebSearch confirms - Hanssen et al., Cell Metabolism]
- Claim: "5-15% increase in metabolic rate"
  [WebSearch confirms - multiple sources]
- All key claims verified.

PHASE 4: DIAGRAMS
- Concept 1: BAT activation mechanism (pathway diagram)
  - Generating SVG...
  - Validating: Text at y=350 in 400-height viewBox, with 40px padding = OK
  - Container at height 120 with 3 text lines (75px) + 80px padding = needs 155px, FAIL
  - Regenerating with taller container...
  - Re-validating: Container now 180px, text fits with padding. PASS.
- Concept 2: Cold vs thermoneutral comparison
  - Generating SVG...
  - Validating: All checks pass.

PHASE 5: ASSEMBLY
- Composing email HTML with diagrams embedded...

PHASE 6: PUBLISH
- Saving to content/social-content/newsletters/2026-01-18-cold-exposure-bat.html
- Updating progress.txt with learnings
- Git commit

Example: Regulatory Brief Run

User: /ralph-content "FDA safety brief for DSIP peptide" --type regulatory_brief

You (thinking through each phase):

PHASE 0: CONTEXT
- Reading progress.txt... Previous regulatory brief run had citation attribution error.
  I'll verify each citation's author/year against PubMed.

PHASE 1: RESEARCH (Extended for regulatory)
- Query 1: "DSIP delta sleep inducing peptide clinical trials safety human studies PubMed"
  [WebSearch returns 4 clinical studies from 1980s-1990s]
- Query 2: "DSIP regulatory approval Russia Deltaran Ministry of Health"
  [WebSearch returns registration info]
- Query 3: "DSIP LD50 toxicity animal studies lethal dose"
  [WebSearch confirms no lethal dose established]
- Evaluating: I have human trial data (n=~200 total), Russian approval, toxicity profile.
  Sufficient for regulatory brief.

PHASE 2: DRAFT + CRITIQUE
- Writing with regulatory brief structure...
- Self-critique against 12-point regulatory rubric:
  ✓ Clinical data with n, design, findings
  ✓ Route-specific table
  ✓ Safety signals categorized (Known/Theoretical/Absent)
  ✓ Regulatory precedent documented
  ✗ Conclusion needs explicit evidence gap statement
- Revising conclusion to explicitly state limitations...
- Re-evaluating: All 12 criteria now pass.

PHASE 3: VERIFICATION (MANDATORY for regulatory_brief)
- Step 3A: Source Credibility Audit
  - Found cosmicnootropic.com in research notes - REJECT (vendor site)
  - Searching for peer-reviewed alternative...
  - Found Popovich et al. 2003 (PMID 12782416) - ACCEPT
- Step 3B: Citation Accuracy Validation
  - Ref [1] claims "Schneider-Helmert 1988" but PubMed 1299794 shows "Bes F et al. 1992"
  - CORRECTING: Update author and year
- Step 3C: Claim Verification
  - Claim: "LD50 never determined" - Verified via Graf & Kastin 1984 review
  - Claim: "97% opiate addicts improved" - Verified via Dick 1984 (PMID 6548969)
- Step 3D: Reasoning Consistency
  - N/A for single document (check if creating related briefs)

PHASE 4: DIAGRAMS
- Regulatory briefs use tables, not diagrams. Skipping.

PHASE 5: ASSEMBLY
- Composing print-optimized HTML with NGM styling...

PHASE 6: PUBLISH
- Saving to content/social-content/newsletters/2026-01-27-fda-safety-brief-dsip.html
- Creating JSON metadata file
- Updating progress.txt with citation validation learning
- Git commit

Quality Over Speed

This skill prioritizes quality over speed. It's acceptable to:

• Run 4-5 research queries if needed
• Iterate on content 3+ times
• Regenerate diagrams multiple times
• Take the time to verify claims

The goal is content that meets Every.to editorial standards—content the team would be proud to publish.

JSON Schema for Content Pipeline

CRITICAL: All content must be saved as JSON files to appear in /content-pipeline. The content-pipeline API reads JSON files from content directories.

LinkedIn Post JSON Schema

{
  "id": "unique-id",
  "createdAt": "2026-01-22T17:30:00.000Z",
  "content": "The full post text with\n\nline breaks preserved...",
  "meta": {
    "alphaIdea": "Short summary used as title in content pipeline",
    "hookType": "curiosity|stakes|contrarian|pattern",
    "wordCount": 195,
    "targetAudience": "longevity medicine professionals"
  },
  "quality": {
    "iterations": 1,
    "passed": true,
    "scores": {
      "pattern_interrupt": true,
      "hook_under_150_chars": true,
      "creates_curiosity": true,
      "has_clear_thesis": true,
      "uses_line_breaks": true,
      "avoids_wall_of_text": true,
      "follows_hook_expand_close": true,
      "has_specific_numbers": true,
      "avoids_jargon": true,
      "has_original_insight": true,
      "stays_focused": true,
      "has_subtle_cta": true,
      "paragraphs_punchy": true,
      "intellectually_honest": true
    }
  },
  "status": "draft",
  "images": []
}

Required fields for display:

• id - Unique identifier
• createdAt - ISO 8601 timestamp (used for sorting)
• content - Full post text
• meta.alphaIdea - Used as title in content pipeline
• quality.passed - Boolean for quality badge
• quality.iterations - Number for iteration count
• quality.scores - Object with individual rubric scores

Newsletter JSON Schema

{
  "id": "unique-id",
  "createdAt": "2026-01-22T17:30:00.000Z",
  "title": "The Newsletter Title",
  "subtitle": "Optional subtitle shown in preview",
  "textContent": "## Full markdown content\n\nWith all sections...",
  "hasHtmlContent": true,
  "status": "draft",
  "meta": {
    "format": "research_synthesis|ai_in_clinic_playbook|deep_dive",
    "length": "short|medium|long",
    "wordCount": 680,
    "estimatedReadTime": 3,
    "targetAudience": "longevity medicine professionals"
  },
  "quality": {
    "iterations": 1,
    "passed": true,
    "scores": {
      "specificity_check": true,
      "evidence_cascade_present": true,
      "colon_technique_used": true,
      "em_dash_clarity": true,
      "one_sentence_emphasis": true,
      "concrete_over_abstract": true,
      "voice_authenticity": true,
      "forward_looking_conclusion": true
    }
  }
}

Required fields for display:

• id - Unique identifier
• createdAt - ISO 8601 timestamp
• title - Used as title in content pipeline
• textContent - Full markdown content (shown in Markdown view)
• hasHtmlContent - Set to true if HTML file exists (enables preview iframe)
• quality.passed, quality.iterations - For quality badge

Note: The HTML file must have the same base filename as the JSON for the preview iframe to work.

Lead Magnet JSON Schema

The content pipeline supports two formats. Use the new format for all new content.

New Format (Recommended)

{
  "id": "unique-slug",
  "createdAt": "2026-01-22T17:30:00.000Z",
  "title": "The Lead Magnet Title",
  "subtitle": "Optional subtitle",
  "slug": "unique-slug",
  "sections": [
    {
      "title": "Section Title",
      "content": ["paragraph 1", "paragraph 2"]
    }
  ],
  "unexpectedDiscoveries": [
    "Discovery 1",
    "Discovery 2"
  ],
  "frameworks": [
    {
      "name": "Framework Name",
      "description": "Framework description"
    }
  ],
  "references": [
    {
      "title": "Reference title with authors and journal"
    }
  ],
  "accessKeyword": "OPTIONAL_KEYWORD"
}

Legacy Format (Still Supported)

{
  "id": "unique-id",
  "title": "The Lead Magnet Title",
  "slug": "unique-slug",
  "created_at": "2026-01-22T00:00:00.000Z",
  "keyword": "KEYWORD",
  "key_findings": [
    { "finding": "Finding text", "source": "Source citation" }
  ],
  "mechanisms": [
    { "mechanism": "Mechanism name", "clinical_takeaway": "Clinical takeaway" }
  ],
  "references": ["Reference 1 as string", "Reference 2 as string"]
}

Required fields for display:

• id - Unique identifier
• createdAt OR created_at - ISO 8601 timestamp
• title - Used as title in content pipeline

Content fields (use one set):

• New: sections, frameworks, unexpectedDiscoveries
• Legacy: key_findings, mechanisms

References: Supports both [{title: "..."}] and ["string"] formats

Keyword: Supports both accessKeyword and keyword

Note: The "View HTML Lead Magnet" button always appears for lead magnets (HTML file must exist with same base name as JSON).

Diagram PDF Export: Lead magnets support diagram-only PDF export via the "Download Diagram PDF" button. Features:

• Cover page with scroll-stopping hook (auto-generated from title), Newsreader/serif typography, and NGM branding
• Access keyword displayed prominently with "Comment below to get the full analysis" CTA—auto-generated from title if not provided in JSON (accessKeyword or keyword field)
• One diagram per page, scaled to maximize page real estate
• JPEG compression for small file sizes (~300-400KB)
• LinkedIn carousel optimized: landscape orientation, swipe hint, visual hierarchy

The hook is auto-generated using pattern matching on the title/subtitle (e.g., "What 30+ Top Researchers Agree On" for consensus topics, "The Shift Nobody Saw Coming" for revolution topics).

Instagram Script JSON Schema

{
  "id": "unique-id",
  "createdAt": "2026-01-22T17:30:00.000Z",
  "meta": {
    "topic": "Short topic used as title"
  },
  "script": {
    "hook": "First 3 seconds text",
    "body": "Main content of the script...",
    "cta": "Call to action text",
    "totalDuration": 45
  },
  "quality": {
    "passed": true
  }
}

Required fields for display:

• id - Unique identifier
• createdAt - ISO 8601 timestamp
• meta.topic - Used as title in content pipeline
• script.hook, script.body, script.cta, script.totalDuration
• quality.passed - Boolean for quality badge

Files

Output Locations

• Newsletters: content/social-content/newsletters/ (both .html AND .json)
• Lead Magnets: content/learn-platform/lead-magnets/ (both .html AND .json)
• LinkedIn: content/social-content/linkedin-posts/ (.json only)
• Instagram: content/social-content/instagram-scripts/ (.json only)

API Endpoints

• View Lead Magnet HTML: GET /api/lead-magnet-html/[slug]
• Download Diagram PDF: GET /api/lead-magnet-diagrams-pdf/[slug] - Generates LinkedIn-optimized PDF with cover page + diagrams (uses puppeteer + jspdf)

Learning Persistence

• Progress: .ralph-content/progress.txt

Context (Read at Start)

• context/every-voice-patterns.md - Voice patterns and techniques
• context/quality-rubrics.md - All quality criteria
• context/diagram-guidelines.md - SVG generation rules
• context/ngm-style-guide.md - Brand and HTML templates