Version: 3.5.0 Platform: Claude Code & Cowork Status: Production-Ready
Transform content requirements into publication-ready, fact-checked, brand-compliant, SEO-optimized content in 20-30 minutes through a 13-agent autonomous pipeline with 19 skills and 10 industry knowledge packs. New in v3.5: Pipeline performance tracking with actual wall-clock timing per phase and token usage estimation. Multi-backend I/O — choose Google Sheets + Drive, Airtable, or local filesystem for tracking and delivery. Backend migration with
/cf:switch-backend. Brand setup Step G asks users to choose their tracking backend. v3.4: Industry Knowledge Packs for subject matter expertise with SME calibration and domain-specific validation. Brand setup auto-generates key files. v3.2/v3.3: Visual Asset Annotator (Phase 3.5) with auto-generated matplotlib charts, structured internal linking markers, Google Sheets tracking, and Google Drive delivery. v3.0: Social adaptation, CMS publishing, content briefs, A/B variants, translation, video scripts, content audits, calendars, style guides, analytics dashboards, connector discovery, and upgraded agents.
- Pipeline Performance Tracking — Actual wall-clock timing per phase (no more placeholder estimates). Token usage estimation with content tokens, agent instruction tokens, and configurable overhead multiplier. Phase 8 completion summary shows real timing table with benchmark comparison
- Multi-Backend I/O — Choose your tracking and delivery backend: Google Sheets + Drive, Airtable, or local filesystem. Airtable handles both tracking and file delivery via record attachments (~2 min setup vs Google's ~5 min)
- Backend Migration —
/cf:switch-backendto change backends anytime with optional data + file migration. Source data is never deleted. Migration is idempotent and resumable - Brand Setup Step G — During brand onboarding, users choose their preferred tracking backend. Google and Airtable are the primary options; local only if explicitly chosen or skipped
- Industry Knowledge Packs — 10 domain-specific knowledge configs (pharma, BFSI, healthcare, legal, real estate, technology, B2B SaaS, e-commerce, consumer goods, education) that make the Content Drafter write as a subject matter expert and the Scientific Validator perform domain-specific checks
- SME Calibration (Phase 3) — Before writing, the drafter loads the industry knowledge pack and calibrates expertise stance, writing conventions, terminology depth, regulatory awareness, evidence standards, and quality signals
- Domain-Specific Validation (Phase 4) — Validates terminology accuracy, evidence standard compliance, regulatory compliance, common pitfalls, and expert quality signals against the knowledge pack
- Brand Key File Auto-Generation —
/brand-setupcan now create brand-profile.json, guardrails.json, and reference-content.md automatically by analyzing the brand's website, existing files, and user input
- Visual Asset Annotator (Phase 3.5) — New agent that generates matplotlib charts from verified research data and creates structured annotation markers for screenshots, diagrams, and images
- Structured Internal Linking (Phase 6) — SEO agent produces machine-readable
<!-- INTERNAL-LINK -->markers that Phase 8 converts to clickable hyperlinks - Google Sheets Tracking — Service account script (
sheets-tracker.py) for content tracking without MCP dependency - Google Drive Delivery — Service account script (
drive-uploader.py) with auto-created folder hierarchies - Self-Service Google Setup — Brand setup walks users through service account creation and configuration
- Drive Knowledge Vault Verification — Brand setup verifies brand files exist in the correct Drive folder structure
- 📱 Social Adaptation — Transform articles into LinkedIn, Twitter/X, Instagram, Facebook, Threads posts with platform-specific formatting
- 🌐 CMS Publishing — Push content to Webflow and WordPress directly via MCP, with HTML export fallback
- 🌍 Translation — Translate content preserving brand voice across 15+ languages with 3 localization levels
- 🎬 Video Scripts — Generate timestamped scripts for YouTube, TikTok, Instagram Reels, and explainers
- 🔬 Content Briefs — Generate research-backed briefs with keyword analysis, competitor gaps, and outlines
- 📊 Analytics Dashboard — Track quality scores over time, phase timing, brand patterns, and trend analysis
- 🔀 A/B Variants — Generate 3-10 headline, hook, and CTA variants with composite scoring
- 📋 Content Audits — Score content freshness (0-100), identify decay, find coverage gaps
- 📅 Content Calendar — Plan production schedules with deadline tracking and Google Calendar sync
- 🎨 Style Guides — Import brand voice from documents/URLs, generate brand profile JSON
- 📄 Custom Templates — Create content type templates beyond the 5 built-in types
- 🔌 Connector Discovery —
/cf:integrationsdashboard and/cf:connectguided setup for 22 connectors - 🤖 AI Overview Optimization — SEO optimizer now structures content for Google AI Overviews and Perplexity
- 📈 Comparative Scoring — Reviewer shows percentile ranking against brand's historical content
- 🎭 Personality Profiles — Humanizer supports 4 configurable profiles: authoritative, conversational, technical, witty
- 🏭 Industry Patterns — Humanizer removes industry-specific AI telltale phrases for healthcare, finance, tech, legal, education
ContentForge is an enterprise-grade content generation system that replaces 6-8 person content workflows with a coordinated multi-agent AI pipeline. Unlike single-prompt tools, ContentForge runs content through 10 specialized quality gates with three-layer fact verification, preventing hallucinations and ensuring brand compliance.
Target Users:
- Digital marketing agencies managing 50-200 brands
- In-house marketing teams with high content volume
- Content operations in regulated industries (Pharma, BFSI, Healthcare, Legal)
- Enterprise brands requiring consistent quality at scale
What Makes ContentForge Different:
- Zero Hallucinations: Three-layer verification (Phases 2, 4, 7) catches fabricated data
- 95%+ Citation Accuracy: All claims traceable to verified sources
- Brand Voice Consistency: Load and apply brand guidelines automatically
- Natural Language: Phase 6.5 Humanizer removes AI writing patterns with 4 personality profiles
- Quality Transparency: Every piece scored 1-10 across 5 dimensions with comparative benchmarks
- Human Oversight: Content <5.0/10 escalates to review, never auto-publishes
- 19 Skills + 7 Commands: Full content lifecycle — from brief to publish to repurpose, with top commands visible in the Customize panel. All skills include argument-hint autocomplete,
/cf:publishhas execution safety (disable-model-invocation), and 3 key skills have structured evals
Sample Output (Article - "Multi-Agent AI Systems"):
- Processing Time: 28 minutes
- Quality Score: 9.0/10 (Grade A)
- Word Count: 1,855 (Target: 1,500-2,000)
- Citations: 14 sources (92% strongly verified)
- SEO: Primary keyword 1.62% density, all critical placements
- GEO: AI Overview optimized, 3 citeable moments, GEO score 8.8
- Readability: Grade 10.4 (Target: 10-12 for articles)
- Humanization: Zero AI patterns, burstiness 0.72 (natural human rhythm)
- Comparative: 94th percentile vs. brand average
- Loops: Zero (approved on first review)
These 7 commands appear in the Commands section of the Customize sidebar, providing quick access to the most common content workflows:
| Command | What It Does |
|---|---|
/create-content |
Run the full 10-phase pipeline — research, draft, fact-check, humanize, publish |
/content-brief |
Generate research-backed briefs with keyword data, competitor analysis, and outlines |
/social-adapt |
Repurpose articles into posts for LinkedIn, Twitter/X, Instagram, Facebook, Threads |
/publish |
Push content to Webflow or WordPress with preview, verification, and HTML fallback |
/translate |
Translate into 15+ languages preserving brand voice, citations, and SEO |
/brand-setup |
Configure brand voice, terminology, guardrails, tracking backend, and auto-generate key files |
/audit-content |
Audit content library for freshness decay, coverage gaps, and optimization opportunities |
| Skill | Command | Purpose |
|---|---|---|
| Content Pipeline | /contentforge |
Full 10-phase production (20-30 min per piece) |
| Batch Processing | /batch-process |
Process 10-50+ pieces in parallel (4-5x faster) |
| Content Refresh | /content-refresh |
Update old content with current data, preserve SEO |
| Skill | Command | Purpose |
|---|---|---|
| Social Adaptation | /cf:social-adapt |
Article → LinkedIn, Twitter/X, Instagram, Facebook, Threads posts |
| CMS Publishing | /cf:publish |
Push to Webflow/WordPress via MCP or HTML export |
| Skill | Command | Purpose |
|---|---|---|
| A/B Variants | /cf:variants |
Generate 3-10 headline, hook, CTA variations with scoring |
| Analytics Dashboard | /cf:analytics |
Quality trends, timing breakdown, brand performance |
| Skill | Command | Purpose |
|---|---|---|
| Translation | /cf:translate |
Translate preserving brand voice, 15+ languages, 3 levels |
| Video Scripts | /cf:video-script |
Timestamped scripts for YouTube, TikTok, Instagram Reels |
| Skill | Command | Purpose |
|---|---|---|
| Content Brief | /cf:brief |
Research-backed brief with keyword analysis and outline |
| Content Audit | /cf:audit |
Freshness scoring, decay detection, gap analysis |
| Content Calendar | /cf:calendar |
Production scheduling with deadline tracking |
| Style Guide | /cf:style-guide |
Import brand voice, generate brand profile JSON |
| Custom Template | /cf:template |
Create content type templates beyond the 5 built-in |
| Skill | Command | Purpose |
|---|---|---|
| Integrations | /cf:integrations |
Dashboard showing connected vs. available connectors |
| Connect | /cf:connect |
Guided setup for any of 22 supported connectors |
| Add Integration | /cf:add-integration |
Add a custom MCP connector for any API or service |
| Switch Backend | /cf:switch-backend |
Switch tracking backend (local/airtable/google) with optional data migration |
| Skill | Command | Purpose |
|---|---|---|
| Help | /cf:help |
User guide, pipeline overview, skill list, examples, troubleshooting |
Phase 1: Research Agent
↓ Quality Gate 1: 5+ live sources, competitor analysis, differentiated angle
Phase 2: Fact Checker
↓ Quality Gate 2: 80%+ verified claims, zero flagged items, all URLs live
Phase 3: Content Drafter ⬆ UPGRADED (SME Calibration via Industry Knowledge Packs)
↓ Quality Gate 3: Word count ±10%, all sections covered, SME calibration applied
Phase 3.5: Visual Asset Annotator [NEW in v3.2]
↓ Quality Gate 3.5: Chart data verified, annotation fields complete, visual density met
Phase 4: Scientific Validator ⬆ UPGRADED (Domain-Specific Validation)
↓ Quality Gate 4: Zero hallucinations, domain terminology verified, regulatory compliance
Phase 5: Structurer & Proofreader
↓ Quality Gate 5: Zero grammar errors, readability on target, brand compliant
Phase 6: SEO/GEO Optimizer ⬆ UPGRADED (Structured Internal Linking)
↓ Quality Gate 6: Keywords optimized, meta tags ready, GEO score ≥7, internal links mapped
Phase 6.5: Humanizer ⬆ UPGRADED
↓ Quality Gate 6.5: AI patterns removed, personality profile applied, industry patterns cleared
Phase 7: Reviewer ⬆ UPGRADED
↓ Quality Gate 7: Score ≥7.0, comparative ranking, trend analysis, recommendations
Phase 8: Output Manager ⬆ UPGRADED (Visual + Link Integration)
↓ .docx + Medium + Substack + Newsletter + PDF + Social Package + Charts
Post-Pipeline Agents (New in v3.0):
- Agent 10: Social Adapter — Extracts shareworthy moments, generates platform-specific posts
- Agent 11: Translator — Element classification, brand voice mapping, cultural adaptation
Feedback Loops:
- Phase 4 → Phase 3.5 (max 1 iteration): If visual chart data doesn't match verified statistics
- Phase 4 → Phase 3 (max 2 iterations): If hallucinations detected
- Phase 6 → Phase 5 (max 1 iteration): If SEO degrades readability
- Phase 7 → Any phase (max 2 iterations): If dimension scores below threshold
- Total loop limit: 5 iterations before human escalation
ContentForge ships with 7 HTTP connectors that work in both Cowork and Claude Code — Notion (knowledge base), Canva (design), Figma (design), Webflow (publishing), Slack (notifications), Gmail (delivery), and Google Calendar (scheduling).
The plugin works fully WITHOUT any connectors — you can run the complete content pipeline and get output locally. Connectors unlock platform integrations.
New in v3.0: Run /cf:integrations to see which connectors are active and what they unlock. Run /cf:connect <name> for guided setup of any of 22 supported connectors across 12 categories.
Claude Code users who need Google Sheets (requirement intake) and Google Drive (brand knowledge vault) can use the advanced setup:
cp .mcp.json.example .mcp.jsonSee CONNECTORS.md for the full connector reference.
Option A: Claude Marketplace (Recommended)
# Search for "ContentForge" in Claude Code marketplace
claude plugins install contentforgeOption B: Manual Install
# Clone repository
git clone https://github.com/indranilbanerjee/contentforge.git
# Move to Claude plugins directory
# On Mac/Linux:
mv contentforge ~/.claude/plugins/
# On Windows:
mv contentforge %USERPROFILE%\.claude\plugins\# Session startup will show:
# ✓ ContentForge v3.5 loaded — Enterprise content production with zero hallucinations
# /contentforge — Single piece (20-30 min)
# /batch-process — Multiple pieces in parallel (4-5x faster)
# /content-refresh — Update old content with fresh data
# /cf:integrations — See connected integrations
# /cf:social-adapt — Repurpose content for social
# /cf:publish — Push to CMSCopy the example config for full npx server support:
cd ~/.claude/plugins/contentforge
cp .mcp.json.example .mcp.jsonEdit .mcp.json with your credentials. See CONNECTORS.md for setup guides.
In Google Drive, create folder structure:
ContentForge-Knowledge/
├── Brand-Name-1/
│ ├── Brand-Guidelines/
│ │ ├── voice-and-tone.md
│ │ ├── terminology.md
│ │ └── visual-identity.pdf (optional)
│ ├── Reference-Content/
│ │ ├── sample-article-1.md
│ │ └── sample-article-2.md
│ └── Guardrails/
│ ├── prohibited-claims.md
│ └── compliance-requirements.md
Populate brand files using templates from config/brand-registry-template.json
/contentforge "Write a 1500-word article about AI content automation for brand AcmeCorp"Expected: Pipeline executes through all 10 phases, outputs .docx with quality scorecard.
| Phase | Agent | Purpose | Avg Time |
|---|---|---|---|
| 1 | Researcher | SERP analysis, source mining, outline creation | 8 min |
| 2 | Fact Checker | URL verification, claim validation, cross-referencing | 3 min |
| 3 | Content Drafter | First draft with brand voice, citations, and SME calibration | 6 min |
| 3.5 | Visual Asset Annotator | Chart generation, visual markers, asset manifest | 2 min |
| 4 | Scientific Validator | Hallucination detection, domain-specific validation, visual data verification | 2 min |
| 5 | Structurer & Proofreader | Grammar, readability, brand compliance | 3 min |
| 6 | SEO/GEO Optimizer | Keywords, meta tags, AI Overview optimization, internal linking | 2 min |
| 6.5 | Humanizer | AI pattern removal, personality profiles, industry patterns | 2 min |
| 7 | Reviewer | 5-dimension scoring, visual + link quality, comparative analysis | 1 min |
| 8 | Output Manager | .docx with charts, internal links, Medium, Substack, PDF | 1 min |
| 9 | Batch Orchestrator | Parallel pipeline coordination, queue management, progress tracking | Post-pipeline |
| 10 | Social Adapter | Extract shareworthy moments, generate platform-specific posts | Post-pipeline |
| 11 | Translator | Element classification, brand voice mapping, cultural adaptation | Post-pipeline |
Total: ~28 minutes per piece (can vary based on complexity and word count)
Gate Structure:
- Criteria: Specific pass/fail conditions (e.g., "Min 5 live sources")
- Decision: PASS (continue), FAIL (fix and re-run), or LOOP (return to earlier phase)
- Feedback: If FAIL or LOOP, specific issues documented for correction
Gate Enforcement:
- All gates enforced automatically by orchestrator
- No human intervention required unless score <5.0 or loops exceeded
- Loop limits prevent infinite iterations (max 2 per phase type, 5 total)
Single-pass fact-checking misses ~15-20% of hallucinations. ContentForge uses layered verification:
- Phase 2 (Fact Checker): Verifies research sources before drafting begins
- Phase 4 (Scientific Validator): Re-verifies drafted content against verified sources
- Phase 7 (Reviewer): Final audit of factual accuracy as part of holistic quality assessment
Result: 100% factual accuracy in production tests, zero hallucinations in published content.
SHA256 hash-based caching (see utils/brand-cache-manager.md):
- On first run for a brand, process all guideline files
- Calculate SHA256 hash of all source files
- Save processed profile with hash to cache file
- On subsequent runs: if hash matches → load cached profile (<5 seconds)
Performance: First run: 2-5 minutes. Cached runs: <5 seconds (95%+ time savings).
| Dimension | Weight | What It Measures |
|---|---|---|
| Content Quality | 30% | Depth, originality, audience value, structure, completeness |
| Citation Integrity | 25% | Factual accuracy, source quality, formatting, recency, cross-referencing |
| Brand Compliance | 20% | Voice/tone, terminology, guardrails, POV, industry compliance |
| SEO Performance | 15% | Keywords, meta tags, on-page elements, GEO readiness, schema |
| Readability | 10% | Reading level, sentence variety, paragraph structure, scannability, humanization |
| Score Range | Grade | Decision |
|---|---|---|
| 9.0-10.0 | A+ to A | APPROVED — Publish + repurpose aggressively |
| 7.0-8.9 | B- to A- | APPROVED — Publish with optional improvements |
| 5.0-6.9 | C- to C+ | LOOP — Return to weakest phase for fixes |
| <5.0 | D to F | HUMAN REVIEW — Escalate, do not auto-publish |
The Reviewer now provides:
- Percentile ranking against the brand's historical content
- Trend tracking across last 10 pieces (strengths, weaknesses, trajectory)
- Score-based recommendations with cross-skill suggestions (e.g., "Run
/cf:social-adapt— 5 shareworthy moments identified")
The ContentForge Differentiator
Phase 6.5 removes AI writing patterns that make content sound robotic, while preserving all SEO keywords from Phase 6.
1. AI Pattern Removal: Scans for and removes 20+ telltale phrases from config/humanization-patterns.json
2. Sentence Variety (Burstiness): Achieves natural human rhythm:
Target: Short (≤12 words) 20% | Medium (13-25) 50% | Long (26+) 30%
Burstiness Score: ≥0.7 (standard deviation / mean)
3. Personality Profiles (New in v3.0): 4 configurable profiles:
- Authoritative — Data-first, definitive, minimal hedging
- Conversational — Direct address, questions, contractions
- Technical — Jargon preserved, precision, process-oriented
- Witty — Wordplay, unexpected comparisons, self-aware asides
4. Industry-Specific Patterns (New in v3.0): Removes industry-specific AI phrases for:
- Healthcare/Pharma (e.g., "healthcare landscape", "paradigm shift in medicine")
- Finance/BFSI (e.g., "financial landscape", "navigate the complexities")
- Technology/SaaS (e.g., "digital transformation journey", "seamless integration")
- Legal (e.g., "navigate the legal landscape", "legal complexities")
- Education (e.g., "educational landscape", "empowering learners")
5. SEO Preservation Check: Verifies keywords unchanged ±2 occurrences
- AI patterns removed: 12-15 per article
- Burstiness: 0.50 → 0.72 (+44%)
- AI detection scores: <30% (vs. 85-95% before humanization)
- SEO keyword variance: ±1 occurrence (negligible)
Possible Causes: Topic too niche, no search volume, web_search not working
Solutions:
- Broaden topic or use related keyword with search volume
- Check Claude's web_search capability is enabled
- Verify internet connectivity
Solutions:
- Review Phase 7 Quality Scorecard for weakest dimension
- Run
/cf:brief— weak briefs lead to weak content - Run
/cf:style-guide— check brand profile completeness - Lower thresholds temporarily in
config/scoring-thresholds.json
Expected behavior: Humanizer auto-loops to Phase 6 for re-optimization. Second pass usually balances both.
Run /cf:integrations to check status. Run /cf:connect <name> for guided setup.
Single-prompt tools produce content in 30 seconds but with ~15-20% hallucination rate, generic voice, and AI writing patterns. ContentForge takes 28 minutes but delivers 0% hallucinations, consistent brand voice, human-sounding prose, and transparent quality scores.
Yes. ContentForge supports three tracking backends: Google Sheets + Drive, Airtable, and local filesystem. During brand setup (Step G), you choose your preferred backend. Airtable requires only a Personal Access Token (~2 min setup). Local works with zero setup. Run /cf:switch-backend to change backends anytime.
ContentForge is free (open source, MIT license). Claude API costs are ~$0.50-1.50 per article depending on length and model.
5 built-in types: Articles (1,500-2,000 words), Blog Posts (800-1,500), Whitepapers (2,500-5,000), FAQs (600-1,200), Research Papers (4,000-8,000). Use /cf:template to create custom types.
Yes. /batch-process handles 10-50+ pieces simultaneously with 4-5x speedup over sequential processing.
Run /cf:social-adapt with any article. It generates platform-specific posts for LinkedIn, Twitter/X, Instagram, Facebook, and Threads with correct character limits, hashtags, and posting time recommendations.
Yes. /cf:translate supports 15+ languages with 3 localization levels (literal, adapted, transcreated). The Translator Agent maps brand voice characteristics to target language conventions.
- Visual Asset Annotator (Phase 3.5) — auto-generated matplotlib charts from verified data
- Structured internal linking markers in SEO agent (Phase 6)
- Chart embedding and TODO boxes in Output Manager (Phase 8)
- Visual quality and internal linking scoring in Reviewer (Phase 7)
- Google Sheets tracking via service account scripts (
sheets-tracker.py) - Google Drive output delivery via service account scripts (
drive-uploader.py) - Self-service Google setup in brand-setup command
- Drive knowledge vault verification during brand setup
- 10 Industry Knowledge Packs for subject matter expertise (pharma, BFSI, healthcare, legal, real estate, technology, B2B SaaS, e-commerce, consumer goods, education)
- SME Calibration step in Content Drafter (Phase 3, Step 0.3)
- Domain-Specific Validation step in Scientific Validator (Phase 4, Step 5)
- Brand-setup key file auto-generation (Step F) — create brand profile, guardrails, and reference content from website analysis
- Argument-hint autocomplete on all 16 user-invocable skills
- Execution safety on
/cf:publish(disable-model-invocation) - Structured evals on 3 key skills (contentforge, cf-brief, cf-style-guide)
- Fixed cf-help missing
namefield in frontmatter
- Pipeline performance tracking — actual wall-clock timing per phase, token usage estimation
- Multi-backend I/O — Google Sheets + Drive, Airtable, or local filesystem
- Backend migration —
/cf:switch-backendwith optional data + file migration - Brand setup Step G — active backend selection during brand onboarding
- 4 new scripts: pipeline-tracker.py, airtable-tracker.py, local-tracker.py, backend-migrator.py
- All 10 pipeline agents instrumented with timing calls
- API mode (REST API for external integrations)
- Real-time collaboration (multiple agents editing simultaneously)
- Custom agent creation (define your own pipeline phases)
- Advanced analytics with ML-powered optimization recommendations
- Image generation integration (DALL-E, Midjourney via MCP)
- Audio content (podcast scripts, voice-over scripts)
Contributions welcome! See CONTRIBUTING.md for guidelines.
Priority Contribution Areas:
- Additional content type templates (landing pages, email sequences)
- Humanization pattern libraries for non-English languages
- Industry-specific quality rubrics
- Test coverage for individual agents
MIT License — see LICENSE file.
Issues: GitHub Issues Discussions: GitHub Discussions
Created by: Indranil Banerjee Built for: Claude Code & Cowork platforms Powered by: Anthropic Claude
Special Thanks:
- Anthropic for Claude and MCP framework
- Digital marketing agencies who provided feedback during beta testing
- Open source community for MCP server implementations
ContentForge v3.5.0 — 13 agents, 19 skills, 10 industry knowledge packs, zero hallucinations. Transform requirements into publication-ready, domain-expert content in 20-30 minutes with pipeline performance tracking and multi-backend I/O.