workflow-templates-broken-zap-2026-06-08/site/docs/ROADMAP.md

AI Content Generation Roadmap

Actionable roadmap for the Hanzo Studio template site AI content generation pipeline.

Last Updated: 2026-02-04
Status: Phase 1 in progress

---

Overview

This roadmap consolidates items from:

• ai-content-generation-strategy.md - Original strategy
• ai-content-generation-research.md - Industry research
• PRD.md - Product requirements

---

Launch Checklist (One-Time)

Task	Status	Notes
Verify hanzo.ai in Google Search Console	🔲 Todo	URL prefix method; see docs/seo-setup-guide.md
Submit sitemap in GSC (/sitemap-index.xml)	🔲 Todo	One-time manual submission
Register with Bing Webmaster Tools	🔲 Todo	Optional; enables IndexNow for Bing/DuckDuckGo
Set OPENAI_API_KEY secret on Vercel	🔲 Todo	Required for production AI content builds
First production deploy via deploy-site.yml	🔲 Todo	Manual dispatch after secrets are set

---

Phase 1: Foundation ✅ (In Progress)

Goal: Basic AI content generation pipeline with knowledge base

Task	Status	Notes
Create content template prompts (tutorial/showcase/comparison/breakthrough)	✅ Done	../knowledge/prompts/
Add CLI flags to generate-ai.ts (--test, --template, --skip-ai)	✅ Done
Sync tutorials from docs repo	✅ Done	84 tutorials synced
Template selection logic based on metadata	✅ Done	selectContentTemplate()
Tutorial context injection for matching templates	✅ Done	findRelevantTutorial()
Quality validation (word count, steps, FAQs, keywords)	✅ Done	checkContentQuality()
Expand model documentation in knowledge base	🔲 Todo	Extract from tutorials
Add concept documentation (inpainting, controlnet, etc.)	🔲 Todo

Deliverable: Working pipeline that generates quality content for top 50 templates

---

Phase 2: Retrieval Quality

Goal: Improve context retrieval accuracy using research findings

Task	Status	Priority	Notes
Implement contextual chunking for tutorials	🔲 Todo	High	Prepend 50-100 token context summaries (49% retrieval improvement)
Add hybrid search (BM25 + embeddings)	🔲 Todo	High	Better than embeddings-only for technical terms
Implement reranking step	🔲 Todo	Medium	Filter top 150 → top 20 chunks
Add uncertainty prompting to templates	🔲 Todo	High	"If unsure about node names, say so" (52% hallucination reduction)
Source attribution in prompts	🔲 Todo	Medium	"Specify source for each claim" (43% reduction)
Lower temperature for factual content	🔲 Todo	Low	0.3-0.5 instead of 0.7

Research basis: Anthropic Contextual Retrieval, NVIDIA chunking study

---

Phase 3: Quality Assurance

Goal: Production-grade content quality pipeline

Task	Status	Priority	Notes
Human review queue for low-confidence content	🔲 Todo	High	Route quality score <60 to review
Brand voice consistency checks	🔲 Todo	Medium	Check for AI artifacts ("In today's fast-paced...")
Hallucination detection for technical claims	🔲 Todo	High	Cross-reference node names against workflow
Plagiarism/originality scoring	🔲 Todo	Low	Compare against existing content
SEO validation (keyword density, meta length)	🔲 Todo	Medium	Automated Surfer-style checks
Content freshness tracking	🔲 Todo	Low	Flag stale content for regeneration

Target: <20% of content requires human review

---

Phase 4: A/B Testing Infrastructure

Goal: Test content variants to optimize for CTR and conversions

Task	Status	Priority	Notes
URL-based variant routing	🔲 Todo	High	?v=tutorial vs ?v=showcase
Cookie-based variant persistence	🔲 Todo	High	Consistent experience per user
Vercel Edge Middleware for A/B	🔲 Todo	Medium	Zero-latency variant serving
GA4 event tracking for variants	🔲 Todo	High	Track CTR by variant
GSC integration for SEO metrics	🔲 Todo	Medium	Impressions, clicks, position
Experiment dashboard	🔲 Todo	Low	View active tests, results

First test: Tutorial vs Showcase for high-traffic templates

---

Phase 5: Scale & Optimization

Goal: Full catalog coverage with automated quality

Task	Status	Priority	Notes
Generate content for all 200+ templates	🔲 Todo	High	After QA pipeline ready
Automated content refresh schedule	🔲 Todo	Medium	Regenerate when templates update
i18n support (Chinese, Japanese)	🔲 Todo	Low	Use existing localized index files
Performance optimization	🔲 Todo	Medium	Batch API calls, caching
Cost tracking and optimization	🔲 Todo	Low	Monitor OpenAI spend

---

Backlog (Future Considerations)

Item	Notes
RAG with vector database	Milvus/Pinecone for larger knowledge base
LLM fine-tuning on Hanzo Studio content	Brand voice consistency
Automated fact-checking against workflow JSON	Verify node references
User feedback integration	Learn from corrections
Multi-model comparison (GPT-4o vs Claude)	Quality/cost tradeoffs

---

Success Metrics

Metric	Current	Phase 1 Target	Phase 5 Target
Templates with AI content	0	50	200+
Avg quality score	N/A	70+	85+
Human review rate	100%	50%	<20%
Indexed pages	0	50+	200+
Organic traffic	0	1k/mo	5k/mo
Cloud CTR	N/A	2%+	3%+

---

Quick Reference: Key Files

File	Purpose
../scripts/generate-ai.ts	Main generation pipeline
../scripts/sync-tutorials.ts	Tutorial sync from docs repo
../knowledge/prompts/system.md	Base GPT-4o system prompt
../knowledge/prompts/*.md	Content template prompts
../knowledge/tutorials/	Synced tutorial content
../overrides/templates/	Human-edited content

---

Commands

# Sync tutorials from docs repo
pnpm run sync:tutorials

# Test generation (first template, no API)
pnpm run generate:ai:test -- --skip-ai

# Generate for specific template
pnpm run generate:ai -- --template flux --skip-ai

# Full AI generation
OPENAI_API_KEY=xxx pnpm run generate:ai

---

See also:

• ai-content-generation-strategy.md
• ai-content-generation-research.md
• PRD.md