v3.8.0 — Context Optimization: 57% Agent Compression + Safety

What's Changed — Structural Performance Release

Why This Release Matters

ContentForge's 10-phase pipeline loads each agent file entirely into context as a system prompt. With agents at 900-1,600 lines, the context window was exhausting by Phase 7, causing:

• Instruction loss (Claude forgets scoring logic mid-evaluation)
• Inconsistent output (different results on the same content)
• Pipeline stalls (context overflow triggers unexpected behavior)

Agent Compression: 11,503 → 4,957 lines (-57%)

Agent	Before	After
07-reviewer	1,600	378
06-seo-geo-optimizer	1,048	319
04-scientific-validator	1,025	274
08-output-manager	996	433
06.5-humanizer	986	273
03-content-drafter	966	264
05-structurer-proofreader	947	269
11-translator	901	291
10-social-adapter	865	287

What was removed: Verbose examples, duplicate explanations, lengthy rubric descriptions, full-length output templates.

What was preserved: ALL quality gates, scoring formulas (weighted 30/25/20/15/10), decision logic (APPROVED/LOOP/HUMAN REVIEW), error handling, phase transitions, loop enforcement, guardrails validation, industry thresholds, and the mandatory completion card.

Agent Safety: maxTurns on All 13 Agents

Each agent now has a hard limit preventing runaway execution:

• Complex phases (Draft, Batch): 30-50 turns
• Standard phases (Research, SEO, Output): 20-25 turns
• Focused phases (Validate, Review, Humanize): 15 turns

Skill Budget Optimization

• All 19 descriptions trimmed to <130 characters (fits within ~15,500 char discovery budget)
• 4 more execution skills get disable-model-invocation: true (total: 5 protected)

31 files changed. 7,620 lines removed. ~20,000 tokens saved per full pipeline run.

Full changelog: CHANGELOG.md

v3.7.2 — Stability: Timeout Fallbacks, Loop Enforcement, Frontmatter Fixes

What's Fixed — Plugin Stability

Why the plugin gets stuck (root causes addressed)

Timeout fallbacks added to the three most stall-prone phases:

• Phase 1 (Research): 90-second SERP timeout. If web search hangs, proceeds with topic-based outline instead of stalling indefinitely.
• Phase 2 (Fact Check): 10-second per-URL timeout. Skips unresponsive URLs instead of blocking on a single dead link. Warns if >50% timeout.
• Phase 3.5 (Visuals): 5-minute image approval timeout. Auto-skips unapproved images to pending_human instead of halting the pipeline.

Loop enforcement in Phase 7:

• Mandatory loop count check before initiating ANY feedback loop
• Hard stop: max 2 loops from Phase 7, max 5 total across pipeline
• Escalates to human review when limits reached (instead of looping forever)

SessionStart hook: Added timeout 30 wrapper on setup.py to prevent session start from hanging if Python has issues.

Frontmatter fixes (5 skills)

• cf-add-integration, cf-switch-backend, cf-template: missing effort field added
• cf-analytics, cf-integrations: missing argument-hint field added

Known architectural issue (future release)

The 07-reviewer.md agent is 1,590 lines — too large for optimal context window usage. This needs to be split into smaller files with progressive disclosure. Same issue affects 5 other agents (900-1,047 lines each). This is a structural refactor planned for a future release.

10 files changed.

v3.7.1 — User Guidance, Phase Progress, Error Messages

What's Fixed

User Guidance Overhaul

• SessionStart redesigned: numbered Quick Start tells users to set up brand first, then create content, then check help
• brand-setup.md: "Quick Start (5 minutes)" at top — 3 questions only. Detailed setup moved below as "Full Setup"
• Troubleshooting: 6 detailed error explanations + pipeline phase timing table (all 11 phases with duration)

Phase Progress Indicators

Agents now show real-time status during content production:

• [1/10] Phase 1: Research Agent — Starting SERP analysis...
• [3/10] Phase 3: Content Drafter — Writing first draft...
• [7/10] Phase 7: Reviewer — Scoring content...
  • APPROVED: score breakdown + "Proceeding to Phase 8"
  • REVISION NEEDED: weakest dimension + loop target + ETA
  • HUMAN REVIEW: issues + user options

Token Tracking Reframed

Removed "token estimate" from all user-facing output. Replaced with Pipeline Complexity metrics (content words, sources cited, quality loops, phases completed). Claude Code/Cowork users are on subscriptions — token counts give false precision.

9 files changed, 200 insertions.

v3.7.0 — Title Curation Overhaul, Brand Validation, Scoring Fixes

What's Fixed — 31 Audit Findings Addressed

Title Curation Overhaul

• Quick SERP reconnaissance before generating titles — scans top 5 competitor titles
• Content-type-specific angles — blog/article/whitepaper/FAQ/research paper each get tailored frameworks
• Brand personality adaptation — titles adjust for authoritative/conversational/technical/witty brands
• Guardrails validation on titles — prohibited terms and claims checked before presenting
• 60-char SERP limit enforced with character counts shown
• Anti-clickbait check + competitor title context for differentiation

Pre-Flight Brand Validation (NEW)

• Brand completeness check runs before every content production
• Regulated industries (pharma/BFSI/healthcare/legal) get explicit warning on empty guardrails
• Phase 3: validates brand profile after loading, warns on empty guardrails
• Phase 5: empty guardrails report "SKIPPED" (not "PASSED"), trigger scoring penalty

Scoring Consistency

• GEO Readiness clarified as SEO sub-score (not phantom 6th dimension)
• Industry threshold overrides: pharma 8.0, BFSI 7.5, healthcare 8.0, legal 8.0
• Dimension minimums enforced (fail if ANY dimension below minimum)
• Empty guardrails: -1.0 Brand Compliance penalty
• Rounding: 1 decimal place, standard rounding

Performance Tracking

• Per-phase timing columns (U-AE) in tracking sheet
• Token estimate and guardrails status columns
• Pipeline performance section in user-facing output

Brand Profile

• New template fields: visual_identity, content_pillars, competitor_analysis
• Brand setup: 4 new steps — audience personas, competitors, pillars, visual identity

Evals

• 3 new tests (6 total): scoring dimensions, empty guardrails, title curation with brand personality

16 files changed, 549 insertions.

Full changelog: CHANGELOG.md

v3.6.0 — AI Image Generation, Platform Features, Quality Hooks

What's New

AI Image Generation (Optional, Human-in-the-Loop)

ContentForge can now generate feature images, contextual illustrations, and social graphics using AI image generation MCP servers — with every image shown to the user for approval before embedding.

New HTTP connectors (Cowork + Claude Code):

• fal.ai — 1,000+ models including Flux 2, SDXL, upscalers
• Replicate — Thousands of image models

New npx servers (Claude Code):

• Stability AI — Generate, edit, upscale, remove backgrounds (11 tools)
• nanobanana — Gemini-powered generation (free tier via Google AI Studio)
• mcp-imagenate — Multi-provider: Gemini, OpenAI gpt-image, Flux 2

How it works:

1. Phase 3.5 checks if image gen MCP is connected
2. Asks user: "Generate images? (full / feature-only / none)"
3. Generates feature image (1200x630), contextual illustrations, diagrams
4. Every image shown to user for approval before embedding
5. Guardrails: max 5 AI images, no text in images, no real people/logos

Pipeline integration:

• Phase 6 SEO: Uses generated feature image for og:image meta tags
• Phase 8 Output: Embeds AI images in .docx with "Image generated by AI" attribution
• Phase 10 Social: Canva MCP generates platform-specific social graphics

Platform Feature Adoption

• effort frontmatter on all 16 skills — optimizes cost and quality per skill complexity
• ${CLAUDE_PLUGIN_DATA} persistent storage — official directory that survives plugin updates
• SubagentStart hook — auto-injects brand voice into every subagent
• Stop hook — quality gate verifying citations, URLs, word count, brand compliance before completion

Summary

Metric	v3.5.1	v3.6.0
HTTP connectors	7	9
npx servers	16	19
Hooks	2	4
Skills with effort	0	16
Image generation	Charts only	Charts + AI images (opt-in)

29 files changed across agents, skills, hooks, scripts, connectors, and docs.

Full changelog: CHANGELOG.md

v3.5.1 — Title Curation Pipeline Fix

What's Fixed

The content production pipeline was skipping title selection — when a user provided a topic, the system auto-generated a single title and immediately started Phase 1 Research. Users had no opportunity to choose or refine the title that anchors the entire content piece.

The Fix

Mandatory Title Curation step inserted before Phase 1:

1. User provides a topic (not a title)
2. System generates 4-5 SEO-optimized title options with different angles:
   • Benefit-driven
   • How-to / Tactical
   • Data-driven / Stat-led
   • Question-based / Curiosity
   • Contrarian / Unexpected
3. User selects, modifies, or provides their own title
4. Pipeline starts with the confirmed title as the anchor

Files Changed

• commands/create-content.md — Added mandatory Title Curation section; input changed from "Topic or title" to "Topic"
• skills/contentforge/SKILL.md — Interactive and quick mode both include title selection; examples updated
• agents/01-researcher.md — Added Step 0.5 (Title Curation) with explicit user confirmation gate; Steps 1 and 6 now use the confirmed title
• README.md — Pipeline diagram shows Title Curation before Phase 1

Why This Matters

The title anchors the entire content piece — research angle, outline structure, SEO optimization, and reader expectations all flow from it. Skipping user approval meant the pipeline was building on an unvalidated foundation.

Full changelog: CHANGELOG.md

ContentForge v3.0.0 — Complete Modernization

ContentForge v3.0.0 — Complete Modernization

12 agents | 17 skills | Zero hallucinations

Delivers every feature promised in v2.0.0 that was never built, adds connector infrastructure matching Digital Marketing Pro, introduces 5 new content management skills, and upgrades all 4 late-pipeline agents with AI Overview optimization, comparative scoring, personality profiles, and industry-specific humanization.

Highlights

14 New Skills

Category	Skills
Publishing & Social	/cf:social-adapt, /cf:publish
Content Optimization	/cf:variants, /cf:analytics
Multilingual & Video	/cf:translate, /cf:video-script
Content Management	/cf:brief, /cf:audit, /cf:calendar, /cf:style-guide, /cf:template
Connector Discovery	/cf:integrations, /cf:connect

2 New Post-Pipeline Agents

• Social Adapter (#10) — Extracts shareworthy moments, generates platform-specific posts for LinkedIn, Twitter/X, Instagram, Facebook, Threads
• Translator (#11) — Element classification, brand voice mapping across 15+ languages, cultural adaptation

4 Agent Upgrades

• Output Manager — 5 new output formats: Medium, Substack, Newsletter, PDF, Social Package
• SEO/GEO Optimizer — AI Overview optimization, citation-worthiness scoring, GEO score
• Humanizer — 4 personality profiles (authoritative, conversational, technical, witty) + industry-specific AI pattern removal (healthcare, finance, tech, legal, education)
• Reviewer — Comparative scoring (percentile vs. brand history), trend tracking (last 10 pieces), recommendation engine with cross-skill suggestions

Connector Infrastructure

• scripts/connector-status.py — 22 connectors across 12 categories
• scripts/setup.py — Session startup validation
• /cf:integrations — Dashboard showing connected vs. available
• /cf:connect <name> — Guided setup for any connector

Stats

• 38 files changed
• 14,754 lines added
• No breaking changes — all v2.x skills work identically

Upgrade Guide

See UPGRADE-GUIDE.md

Full Changelog

See CHANGELOG.md

v2.1.0 — HTTP Connector Architecture

ContentForge v2.1.0 — HTTP Connector Architecture

What Changed

Rebuilt the MCP integration layer following Anthropic's official plugin pattern.

New Files

• .mcp.json — 6 HTTP connectors: Notion, Canva, Webflow, Slack, Gmail, Google Calendar
• CONNECTORS.md — Connector categories with ~~category placeholder pattern

Fixes

• Agent names normalized to kebab-case for proper Cowork routing
• Removed non-standard skill_type: command from skill frontmatter
• plugin.json stripped to 4 fields matching Anthropic's format
• Un-gitignored .mcp.json (HTTP URLs only, no credentials)

Note on Google Sheets/Drive

Google Sheets and Google Drive do not have HTTP MCP endpoints. Claude Code users who need these can use .mcp.json.example (npx configuration). See CONNECTORS.md.

Full changelog: CHANGELOG.md

v2.0.2 — Cowork Compatibility & Agent Accuracy

fix: Add YAML frontmatter to all 10 agents, fix Output Manager MCP to…

v2.0.1 - Critical Installation Fix

🐛 Critical Patch: Installation Issues Resolved

This patch release fixes critical marketplace validation failures preventing ContentForge from installing properly in Cowork.

Issues Fixed

✅ "Manage Plugin" redirects to core → Now opens plugin management UI correctly
✅ Marketplace asks to install after already installed → Installation state now properly tracked
✅ Plugin validation fails → All schema violations resolved

Root Cause

ContentForge v2.0.0's plugin.json declared 7 skills but only 3 existed as files:

• ✓ contentforge/SKILL.md (existed)
• ✓ batch-process/SKILL.md (existed)
• ✓ content-refresh/SKILL.md (existed)
• ✗ generate-variants/SKILL.md (missing)
• ✗ multilingual-content/SKILL.md (missing)
• ✗ content-analytics/SKILL.md (missing)
• ✗ publish-content/SKILL.md (missing)

This violated Claude Code's marketplace validation rules, causing installation to fail.

Changes in v2.0.1

Fixed:

• Removed invalid skills array from plugin.json (now using auto-discovery)
• Removed non-standard fields (capabilities, requirements, target_users, use_cases, performance)
• Skills now auto-discovered from skills/ directory following official plugin architecture

Added:

• hooks/hooks.json configuration for SessionStart banner and PreToolUse hallucination detection
• Proper plugin structure following official Claude Code plugin reference

Cleaned:

• Removed legacy SKILL.md at root
• Removed backup files (.mcp.json.example.backup)
• Removed temporary release files

Migration from v2.0.0

If you installed v2.0.0, simply update to v2.0.1:

/plugin update contentforge@neels-plugins

No configuration changes needed. All existing functionality preserved.

Technical Notes

• plugin.json now follows official Claude Code schema exactly
• Skills discovered from skills/ directory: contentforge, batch-process, content-refresh
• Hooks configuration added for hallucination detection during content generation
• Version: 2.0.0 → 2.0.1

---

Full Changelog: v2.0.0...v2.0.1