content-management

name	content-management
description	CMO-style autonomous content agent. Observes user's work activity (sessions, git, kanban), curates narrative-worthy material, produces multi-format drafts (blog + X thread), maintains narrative arc continuity, and generates visual assets (SVG cover, Mermaid, Excalidraw PNG). session: "2026-06-14 content-manager-cmo-agent" decision: "Agent profile + cron trigger + narrative_arc.md tracking file (not a multi-stage pipeline). SVG over paid image API (user: '비용 발생안시키고 싶어'). Excalidraw→PNG via Puppeteer for architecture diagrams."
category	autonomous-ai-agents
tags	["agent","content","cmo","publishing","narrative","svg","excalidraw","mermaid"]
created	"2026-06-14T00:00:00.000Z"
updated	"2026-06-14T00:00:00.000Z"
links	["[[agent-profiles]]","[[skills/content-pipeline]]","[[@memory/content/narrative_arc.md]]"]

Content Management — CMO Agent

Autonomous content manager that observes what the user builds and turns it into publishable content with rich visuals. Uses a single agent profile + cron trigger, NOT a multi-stage pipeline.

Architecture

Agent profile: content-manager (deepseek-v4-pro)
Cron: daily at 12:00 KST
Knowledge base: brand-guide + glossary + narrative_arc + content-inventory
All files under: P4-cortex/content/

Per-run output:
  memories/insights/
    ├── YYYY-MM-DD-slug.md                ← Blog post (with Mermaid diagrams inline)
    ├── YYYY-MM-DD-slug-cover.svg          ← SVG cover illustration (1200x630, dark theme)
    ├── YYYY-MM-DD-slug.excalidraw.json    ← Excalidraw architecture diagram
    ├── YYYY-MM-DD-slug.png                ← Excalidraw exported to PNG
    ├── YYYY-MM-DD-slug-thread.txt         ← X thread (10-15 tweets)
    └── YYYY-MM-DD-slug.excalidraw         ← Excalidraw binary (from CLI)
  
  P4-cortex/content/
    ├── narrative_arc.md                   ← Updated with new episode
    └── content-inventory.md               ← Updated with new entry

Data Sources

In addition to direct work activity, content-manager monitors two automated collection pipelines:

1. Trend Harvester

• Location: @memory/growth/trend-harvester/{collected,applied,evaluated}/
• Content: GitHub trending repos scored via 5-axis philosophy filter (6h cycle)
• Use: Pick applied/evaluated items with high scores → write "AI Tool of the Week" posts or deep-dives into tools that got integrated into {{AGENT_NAME}}

2. SEO Article Harvester

• Location: P2-hippocampus/knowledge/seo-articles/YYYY/
• Content: 1,642+ SEO/marketing articles from 28 RSS feeds (6h cycle)
• Use: Surface notable Google updates, agentic web trends, or shifts in AI/discovery landscape → write analysis/opinion posts linking to YOUR_DOMAIN's own ARD integration

Agent Profile

Defined at ~/.{{AGENT_NAME_LOWER}}/agents/content-manager.md. Full instructions include:

1. Load Knowledge Base — Read brand-guide, glossary, content-inventory, narrative_arc
2. Gather Context — session_search, git log, kanban completions, harvester outputs
3. Mine for Stories — Score 1-10 (reader value, drew-angle, narrative fit). ≥7 proceed
4. Draft Content — Blog post with embedded SVG cover + Mermaid + Excalidraw PNG
5. Publish Draft — create_post via WordPress MCP (status=draft)
6. Update Tracking Files — narrative_arc.md + content-inventory.md

Web Research

Content-manager has web toolset enabled. Before drafting, optionally search for:

• Related projects or approaches (what others are doing)
• Technical background to strengthen the narrative
• Memes or cultural references that fit the story
• Related blog posts / discussions

Search results strengthen the draft's credibility and relevance.

Visual Asset Pipeline

Three types of images produced per post, all $0 cost, plus optional meme SVGs:

1. Cover Image (CF Workers AI, $0, 1024×512)

AI-generated conceptual illustration via FLUX on Cloudflare Workers AI. Four scene templates:

• garden — 태양 정원 (Creative)
• city — 덩굴 도시 (Build Log)
• shepherd — 사막의 양치기 (AI & Tools)
• connector — 연결자 (Systems)

All feature: anthropomorphic blue sheep, lion silhouette, geometric sun, cracked earth. See scripts/cover_gen.py for prompts. Featured image via MCP create_post(featured_image=...). Fallback: SVG templates at references/svg-templates/.

1b. Meme SVG (optional — cultural reference)

If the story has a natural meme angle, create a companion SVG. Recognizable meme formats:

Template	Use Case
Drake Reject/Approve	Before/after comparison, old vs new
"This is fine" (burning)	Recognizable pain, debugging stories
Galaxy brain	Escalating understanding, growing insight
Distracted boyfriend	Three-way comparison, options trade-off

See references/svg-meme-templates.md for SVG code templates.

2. Mermaid Diagrams

Inline ````mermaid` code blocks. Quartz renders natively as SVG. Types: graph TD/LR, sequenceDiagram, flowchart

3. Excalidraw → PNG

For complex architecture/flow diagrams. Two-step process:

1. Create .excalidraw.json file with elements array (rectangles, text, arrows, etc.)
2. Convert to PNG via headless Chrome: scripts/excalidraw-to-png.js

NODE_PATH=~/.{{AGENT_NAME_LOWER}}/scripts/node_modules \
  node ~/.{{AGENT_NAME_LOWER}}/scripts/excalidraw-to-png.js \
    input.excalidraw.json \
    output.png

Knowledge Base Files

File	Path	Purpose
brand-guide	P4-cortex/content/brand-guide.md	Voice, audience, platform strategy
glossary	P4-cortex/content/glossary.md	Project terms ({{AGENT_NAME}}, M-LOG, etc.)
narrative_arc	P4-cortex/content/narrative_arc.md	Published content continuity tracking
content-inventory	P4-cortex/content/content-inventory.md	All drafts/published with dedup topics
writing-style-guide	P1-limbic/persona/writing-style-guide.md	Tone, forbidden expressions, Korean voice
form-template	https://dev.to/siiddhantt/building-reefwatch-a-coral-powered-production-triage-agent-23hf	Canonical blog post 구조 — 반드시 이 형태를 따라라

Post-Publish Enrichment — Links & References

After drafting, enrich the post with links before saving as draft:

1. 내부 링크: Scan narrative_arc.md and content-inventory.md for existing posts on related topics. Add internal links where natural (e.g., "v0.8 simplification post ([link])에서 다뤘듯이...")
2. 외부 참조: Check SEO harvester articles for relevant sources (e.g., Google announcements, agentic web trends). Link to original sources where cited.
3. 상호 참조 클러스터: Posts on the same pillar (BUILD LOG, AI & TOOLS, SYSTEMS, CREATIVE) should interlink. Episode 1→2→3 of a season should form a chain.
4. 용어 링크: First mention of {{AGENT_NAME}}/M-LOG/ARD 등 should link to the glossary or relevant explanation if it's the first post in a series.

Format example:

<!-- wp:paragraph -->
<p><strong>덜어내는 작업</strong>은 <a href="/?p=XX">v0.8 architecture compression 포스트</a>에서 자세히 다뤘다. 이번 글은 그 연장선에 있는 <strong>연결하기</strong>에 대한 이야기다.</p>
<!-- /wp:paragraph -->

Voice & Form (Critical — Read Before Drafting)

Form Template (ReefWatch)

이 구조를 반드시 따라라. 모든 포스트는 이 템플릿을 프레임으로 사용한다.

1. Hook-first opening — 짧고 강한 문장. 문제를 직접 던진다.
   "Production incidents almost never break in one place."
   "A normal chatbot... But that is not triage. That is a polished to-do list."

2. Problem → Decision → Build Path
   "I wanted something more useful: an agent that could..."
   "The design constraint from the start: no evidence, no answer."

3. "What This Guide Builds" — bullet list of deliverables
4. Why <technology> section — "Why [X] belongs at the center" — 결정 이유
5. Build path (slices) — numbered slices showing progressive build:
   "Slice 1: Prove [X] Can Be The Data Plane"
   "Slice 2: Keep [X] Warm"
   Structure: What I Built → Why It Mattered
6. Architecture diagrams (Mermaid 또는 SVG)
7. Key design decisions with failure modes table
8. Personal closing ("Thanks for reading...")

Voice Rules

1. 반말 + 1인칭 ("저", "나"). 독자에게 "당신"으로 직접 말한다.
2. 금지: "이 글은", 추상 주어, 존댓말, AI투 카피 ("~에 대해 알아보겠습니다")
3. Bold 로 핵심 문장 강조
4. 짧은 문장. 문제 → 바로 해결.
5. 개인적인 깨달음을 중심으로: "That was useful because it proved..."
6. 거대한 시스템 프롬프트를 거부한다: "I stopped trying to make one heroic system prompt do everything"
7. 숫자를 써라: "14→6 agents", "43→25 scripts"
8. 딱 하나의 주장만: 포스트 하나에 하나의 테제

Before drafting, study https://YOUR_DOMAIN/ `s published posts (ID 12, 13) to absorb Drew's actual voice.

Content Pillars (Editorial Judgment)

1. BUILD LOG — {{AGENT_NAME}} infra, architecture, troubleshooting
2. AI & TOOLS — Agent systems, tool reviews, pattern discovery
3. SYSTEMS — Design philosophy, decision frameworks, taste
4. CREATIVE — M-LOG, side projects, experiments

Pitfalls

• Draft must feel like Drew wrote it, not an AI: If the draft sounds like generic LLM content, rewrite it. If you can't match the voice, produce nothing.
• SILENT is correct: If nothing worth publishing, produce nothing. Don't force output.
• Quality over quantity: One great post per run beats five mediocre ones. The cron runs daily, backlog clears over time.
• Never ask Drew what to write: The CMO decides. Present the draft.
• CF Workers AI FLUX for covers: Free tier, $0. cover_gen.py generates 1024×512 PNG via 4 scene templates. SVG templates as fallback only.
• Excalidraw PNG for architecture: Complex diagrams via headless Chrome export.
• Try first, retreat when proven: User explicitly corrected: "변화를 두려워 하지마. 상상만으로 결정하지마. 상정한 방향을 현재 환경에 최대한 적용해보고, 그래도 비효율이면 그때 물러서." Apply the best approach to the current environment first; only retreat when the data says it's inefficient — not from imagined objections.

WordPress Publishing Pipeline

Content-manager publishes to YOUR_DOMAIN (WordPress + GeneratePress) via the WordPress MCP server.

MCP server: Registered in ~/.config/opencode/opencode.jsonc as mcp.wordpress. Script: ~/scripts/wordpress-mcp-server.js — wraps wp-cli via docker exec. Tools: create_post, upload_media, list_posts, get_post, create_category, set_site_option, set_theme_mod.

Content Format (Critical)

WordPress uses the Gutenberg block editor. Post content MUST be Gutenberg-compatible HTML, NOT raw Markdown.

Good:

<!-- wp:paragraph -->
<p><strong>핵심 문장은 Bold로.</strong> 이렇게 써야 제대로 렌더링된다.</p>
<!-- /wp:paragraph -->

<!-- wp:heading {"level":2} -->
<h2>슬라이스 제목</h2>
<!-- /wp:heading -->

<!-- wp:code -->
<pre class="wp-block-code"><code>code block here</code></pre>
<!-- /wp:code -->

Bad (will show raw text):

## Markdown 제목
**볼드** _이탤릭_

Rules:

• NO YAML frontmatter — WordPress stores metadata separately
• NO raw Markdown (#, ##, **, *, - ) — use HTML tags
• ALL content must be wrapped in Gutenberg comment blocks (<!-- wp:paragraph -->...<!-- /wp:paragraph -->)
• Use <!-- wp:list --> for lists, <!-- wp:code --> for code blocks, <!-- wp:heading {"level":2} --> for headings
• Use <!-- wp:html -->...<!-- /wp:html --> for custom HTML sections (cards, hero, etc.)

Pipeline (2026-07-01 redesign)

content-curator (script, 08:00/15:00, $0)
  → reads: trend keep, SEO articles, git, kanban, narrative arc, backlog
  → heuristic dedup + scoring
  → kanban INSERT (content-write / trend-review / creative-write)
  → office-autopilot (5m) picks up → orchestrator → content-manager agent
    → writes post (Gutenberg HTML, SVG cover)
    → calls create_post(status="publish")  (creative-write만 draft)
    → content-editor (agent, 12:00/20:00) → 사후 QA
    → updates narrative_arc + content-inventory

규칙:

• No CLI approval needed. Pipeline is fully autonomous.
• create_post 호출 시 반드시 slug, author, category 전달. tags는 절대 전달 금지.
• slug: 영어 kebab-case (예: model-routing-architecture). 절대 한글/빈 값 금지.
• author: 1 (YOUR_USERNAME) — 전 카테고리 공통
• category: 카테고리 이름(ID 아님): "Build Log", "AI & Tools", "Systems", "Creative"
• Creative pillar는 explicit-only. 자동 감지 금지. kanban task / 사용자 요청 / creative-backlog.md 참조 시에만 작성.
• SVG cover를 생성했다면 featured_image에 절대 경로 전달
• Draft 생성 후 editor agent가 자동 QA → 수정 → publish
• 사용자가 마음에 안 들면 wp-admin에서 unpublish → 다음 run 때 반영
• "이건 publish하지 마" 같은 피드백은 memory에 저장되어 다음 draft에 반영됨

Site Config

Property	Value
URL	https://YOUR_DOMAIN
Admin	https://YOUR_DOMAIN/wp-admin
Theme	GeneratePress 3.6.1
Fonts	Noto Sans KR (body), Noto Serif KR (headings), JetBrains Mono (code)
Colors	#fafaf8 bg, #8b7355 accent, #1c1c1a text
Layout	no-sidebar, separate-containers
Tunnel	Cloudflare Tunnel (launchd: ai.{{AGENT_NAME_LOWER}}.cloudflared-wp)

Relationship to content-pipeline Skill

This skill supersedes the older content-pipeline skill which was based on trend-harvester/SEO sources with a 3-hour cycle. The new approach:

• Source: User's own work (sessions, git, kanban) vs external trends/SEO
• Cycle: Daily, material-driven vs fixed 3-hour
• Agent: Single autonomous profile vs multi-stage kanban pipeline
• Images: SVG + Excalidraw PNG + Mermaid vs text-only drafts

The two overlap. Consolidation deferred to background curator.