AI Directive Integration

bloxx cms — 10 features · 4 phases

From static content blocks to an AI-powered system where blog posts stay live, suggestions arrive as you type, and Claude generates CMS-connected content automatically.

Foundation — Built

Shipped Infrastructure

Directive System
6 types: @faq @team @testimonials @place @reviews @url
DirectiveBlock
TipTap node with pending/resolved/stale/error states
Batch Resolver
POST /resolve, photo proxy, design tokens
RenderKit Registry
Variant auto-selection, schema.org JSON-LD
Blog Generation
claude-sonnet-4-6 with tool_use
MCP Intel Tools
Citation health, competitor snapshots, page audits
Schema Registry
130+ types, vertical recommendations
Scheduled Workers
Citation checks, competitor crawl, link health
@url Dual-Layer
Direct fetch + Firecrawl fallback
Phase 1 — Foundation

01 Schema Export Planned

Every downstream AI feature consumes this contract. Add a new directive and AI knows about it automatically. Single dependency for the entire roadmap.

Phase 2 — AI Integration

02 AI Emits Directives Planned

Blog posts stay live — when FAQs change in the CMS, the generated post updates too. Claude emits @faq --limit=5 instead of baking in static HTML.

Key decision: Prompt-based first (streaming prose). Graduate to tool_use only if syntax errors >5%.

03 Tier 1 Suggestions Planned

Zero-cost, zero-latency suggestions via regex. Type “our team” and a sidebar chip offers @team. Tab to accept. Works on free tier — first taste of AI assistance.

04 @ai Meta-Directive Planned

Any user prompt becomes a context-aware HTML block. @ai "comparison table of our plans" resolves through Claude with full site context.

Key decision: Nested directives deferred — @ai emitting @faq creates a resolution dependency chain that’s not solved yet.

05 Discoverability Score Planned

Users see WHY directives matter. A toolbar indicator shows schema.org coverage building up as they add directives: “3/7 recommended schemas active.”

Phase 3 — Advanced AI

06 Intel-Driven Suggestions Planned

“Competitors all have FAQ schema” is 10x more compelling than “you mentioned questions.” Suggestions powered by real competitive intelligence data.

Key decision: Max 3 per session, back off after dismissals. Suggestion fatigue kills adoption.

07 Schema as Tool Definitions Planned

Each directive becomes a Claude tool automatically. Structured output with validated arguments — no per-directive prompt engineering needed.

08 Full Loop Planned

Click “generate blog post” → complete article with live CMS data. FAQ, reviews, team bios all resolve from real content.

Key decision: Draft/accept flow — show placeholders first, resolve only after user approves. Saves API calls on content they might rearrange.
Flywheel — Vision

09 Autopilot Recommendations Vision

Closed loop — weekly citation health scans generate prioritized directive recommendations surfaced in an Action Queue dashboard.

10 @url Browser Rendering Vision

Cloudflare Browser Rendering replaces Firecrawl for JS-heavy URL previews. Native, no external dependency, with product schema extraction in Phase 2.

Dependency Chain
flowchart TD
  F["Foundation (Built)"]
  S1["01 Schema Export"]
  S2["02 AI Emits Directives"]
  S3["03 Tier 1 Suggestions"]
  S4["04 @ai Meta-Directive"]
  S5["05 Discoverability Score"]
  S6["06 Intel Suggestions"]
  S7["07 Schema as Tools"]
  S8["08 Full Loop"]
  S9["09 Autopilot Recs"]
  S10["10 @url Upgrade"]

  F --> S1
  S1 --> S2
  S1 --> S3
  S1 --> S4
  S1 --> S5
  S2 --> S7
  S3 --> S6
  S4 --> S7
  S7 --> S8
  S6 -.-> S9
  S8 -.-> S9
  F -.-> S10

  classDef built fill:#2d6a4f33,stroke:#2d6a4f,stroke-width:2px
  classDef planned fill:#2563a033,stroke:#2563a0,stroke-width:1.5px
  classDef vision fill:#8b9aab22,stroke:#8b9aab,stroke-width:1px,stroke-dasharray:4

  class F built
  class S1,S2,S3,S4,S5,S6,S7,S8 planned
  class S9,S10 vision
  
Design Decisions
01
Prompt-based emission first
Streaming prose doesn’t pair with tool_use stop-and-resume. Graduate to tools only if >5% syntax errors.
02
Nested @ai directives deferred
Resolution dependency chains need UX work. Ship @ai as self-contained HTML snapshots first.
03
Draft/accept before resolution
Don’t resolve immediately — let users rearrange, then batch-resolve on approval.
04
Max 3 Tier 2 suggestions/session
Suggestion fatigue kills adoption faster than missing suggestions. Back off after dismissals.
05
Placement guidelines in prompt
Social proof after claims, @faq near end, @cta after value prop. Baked into generation context.
06
Cloudflare /scrape for @url MVP
Native browser rendering, no external Firecrawl dependency. Schema extraction is Phase 2.