agent-content-pipeline（Agent 内容管道）

Content Pipeline Skill

Safe content automation with human-in-the-loop approval. Draft → Review → Approve → Post.

Setup

npm install -g agent-content-pipeline
content init . # Creates folders + global config (in current directory)

For cryptographic approval signatures (password-protected):

content init . --secure

This creates:

• drafts/ — work in progress (one post per file)
• reviewed/ — human reviewed, awaiting your revision
• revised/ — you revised, ready for another look
• approved/ — human-approved, ready to post
• posted/ — archive after posting
• templates/ — review and customize before use
• .content-pipeline/threads/ — feedback thread logs (not posted)

Your Permissions

✅ Can do:

• Write to drafts/
• Read all content directories
• Revise drafts based on feedback
• Move revised files to revised/
• Run content list to see pending content

❌ Cannot do:

• Move files to approved/ (only the human can approve)
• Post content
• Set status: approved

Creating Content

One post per file. Each suggestion or draft should be a single post, not a collection.

File naming: YYYY-MM-DD-<platform>-<slug>.md

Use frontmatter:

---
platform: linkedin    # linkedin | x | reddit (experimental)
title: Optional Title
status: draft
subreddit: programming  # Required for Reddit
---

Your content here.

Tell the human: "Draft ready for review: content review <filename>"

The Review Loop

drafts/ → reviewed/ → revised/ → approved/ → posted/
              ↑          │
              └──────────┘
               more feedback

1. You write draft to drafts/
2. Human runs content review <file>:

- With feedback → file moves to reviewed/, you get notified - No feedback → human is asked "Approve?" → moves to approved/

1. If feedback: you revise and move to revised/
2. Human reviews from revised/:

- More feedback → back to reviewed/ - Approve → moves to approved/

1. Posting happens manually via content post

After Receiving Feedback

When you get review feedback:

1. Read the file from reviewed/
2. Apply the feedback
3. Move the file to revised/
4. Confirm what you changed
5. (Optional) Add a note: content thread <file> --from agent

Platform Guidelines

LinkedIn

• Professional but human
• Idiomatic language (Dutch for NL audiences, don't be stiff)
• 1-3 paragraphs ideal
• End with question or CTA
• 3-5 hashtags at end

X (Twitter)

• 280 chars per tweet (unless paid account)
• Punchy, direct
• 1-2 hashtags max
• Use threads sparingly
• If Firefox auth fails, you can paste auth_token and ct0 manually

Manual cookie steps: 1) Open x.com and log in 2) Open DevTools → Application/Storage → Cookies → https://x.com 3) Copy auth_token and ct0

Reddit (experimental)

• Treat as experimental; API and subreddit rules can change
• Requires subreddit: in frontmatter
• Title comes from frontmatter title: (or first line if missing)
• Match each subreddit's rules and tone

Commands Reference

content list                    # Show drafts and approved
content review <file>           # Review: feedback OR approve
content mv <dest> <file>        # Move file to drafts/reviewed/revised/approved/posted
content edit <file>             # Open in editor ($EDITOR or code)
content post <file>             # Post (prompts for confirmation)
content post <file> --dry-run   # Preview without posting
content thread <file>           # Add a note to the feedback thread

Security Model

The security model separates drafting (AI) from approval/posting (human):

• ✅ Agent drafts content
• ✅ Agent revises based on feedback
• ❌ Agent cannot approve (human approves via content review)
• ❌ Agent cannot post

Posting is handled manually via CLI — never by the agent directly.

Platform-specific security

Both platforms require password to post. Tokens are extracted from Firefox and encrypted locally.