Build an AI Social Media Content Pipeline: Coach, Writer, Grader, Scheduler

Posting consistently is the homework assignment nobody enjoys. You know the routine: stare at the blank composer, write something in a hurry, manually resize the idea for three platforms, paste it into each app, second-guess the wording, and repeat tomorrow. Most people don't fail at social media because they lack ideas — they fail because distribution eats every spare hour.

The fix is not "use AI to write your posts." That produces generic filler at scale. The fix is a pipeline: one long-form idea enters, and platform-tailored, quality-checked, scheduled posts come out — with you approving rather than typing. Here's how to build it, and why the pieces are simpler than they sound.

The three layers

The critical insight: the logic lives in text files you own. The model is swappable. That portability is what keeps you from being locked into any single AI vendor.

Layer one: teach the AI who you are

Before any post gets written, the AI needs your brand brief — a short document that answers three questions: who is the audience, what problem do you solve, and what does your voice sound like. Written once, prepended to everything:

# brand_brief.md
- Audience: solo operators and small teams evaluating AI automation
- Problem we solve: too many AI "tips", too few working systems
- Voice: plain English, concrete examples, light humor, zero hype words
- Banned: "revolutionary", "game-changer", "10x your income", engagement-bait questions
- Formats: LinkedIn (opinion + lesson), X (single idea, <200 chars), IG (carousels with captions)

With that in place, drafts stop sounding like a random chatbot and start sounding like you on a good day.

Layer two: skills, not prompts

A prompt is a one-off request. A skill is a reusable instruction set with rules, constraints, and a defined output shape. A working pipeline needs about five:

• Content Coach — picks what's worth posting this week from your topics and past performance.
• Post Writer — drafts for one specific platform at a time, following the brand brief.
• Viral Hooks — a library of proven first lines; the hook decides whether anything else gets read.
• Post Grader — quality control: the AI scores its own draft against your rubric before you see it.
• Repurpose Engine — turns one long piece into a week of platform-specific posts.

The grader deserves special attention, because it's the difference between automation and embarrassment. It works from an explicit rubric:

{
  "rubric": {
    "hook": "First line creates curiosity or names a concrete problem (0-3 pts)",
    "specificity": "Contains at least one concrete example, number, or named tool (0-3)",
    "voice": "Matches brand brief; no banned words (0-3)",
    "format": "Respects platform length and structure rules (0-3)"
  },
  "pass_threshold": 10,
  "on_fail": "return draft with fixes applied, then re-grade once"
}

A draft scoring 8 out of 12 never reaches your queue. This one step removes most of the "AI slop" problem — the model argues with itself so you don't have to.

Building the hook library

The Viral Hooks skill is the cheapest upgrade with the biggest effect, because the first line decides whether the other nineteen lines get read. Build it like a swipe file: every time a post performs, capture why the opening worked and file the pattern. Ten to twenty entries is plenty, and each one is a single line with a note:

# hooks.md
- "I spent [N] hours doing X manually. Here's the system that replaced it." (specific cost → curiosity)
- "Everyone recommends [popular tool]. I stopped using it because [concrete reason]." (mild heresy, defensible)
- "The mistake I see in every [category] workflow: [named mistake]." (targets a recognizable pain)
- "[Number] signals your [process] is ready to automate." (list promise, scannable)

The writer skill draws from this file instead of inventing openers from scratch — which is exactly where unaided models default to "In today's fast-paced world…". A hook library is also honest: every entry is a pattern you personally observed working, not borrowed virality from someone else's audience.

Layer three: the publishing bridge

Models can write; they cannot post. That last mile belongs to middleware. In an n8n workflow, the flow is explicit:

# n8n: graded draft → scheduled post
trigger:
  type: schedule
  cron: "0 9 * * 1"        # Monday 9 AM planning run
steps:
  - source:  { service: youtube, action: get_transcript, video: latest }
  - repurpose:
      model: qwen2.5:7b
      url: "http://ollama:11434/api/chat"
      outputs: [linkedin_post, x_thread, instagram_carousel]
  - grade:   { rubric: brand_rules.json, threshold: 10, retry: 1 }
  - queue:   { service: postiz, schedule: "next free slot", platforms: [linkedin, x, instagram] }
  - notify:  { channel: email, message: "3 posts queued — approve by 6 PM" }

One transcript in, three graded posts queued — with a human approval gate before anything goes live, which is how you keep autonomy safe while it earns trust.

The repurposing math

The economics of the whole pipeline come from one asymmetry: creating a new idea is expensive, reformatting an existing one is nearly free.

1. Publish the hero piece — a video, a long article, a podcast episode.
2. The Repurpose Engine extracts the key points and drafts platform variants: three LinkedIn posts, a week of X posts, carousel scripts.
3. The Post Grader filters and polishes.
4. The scheduler spaces them across the coming week.

One recording session becomes a week of presence. This very blog you're reading runs on the same pattern — a scheduled job picks a topic, generates the images, publishes the post, and queues the social shares, with the creative direction set once by a human.

"But I use ChatGPT/Gemini, not Claude"

Good news: the assets are portable. Skills are markdown; rubrics are JSON. Paste them into a Custom GPT, a Gemini Gem, or hand them to a local model through Ollama. The pipeline's value lives in the structure — brief, skills, grader, scheduler — not in any one model. Swap the intelligence layer whenever something better ships; your system survives every upgrade.

Final thoughts

Automating social media badly means letting an AI write filler at scale. Doing it well means building a small editorial machine: a brand brief that carries your voice, five focused skills that coach, write, and grade, and a scheduler that handles distribution while you sleep. You keep the two jobs that matter — choosing the ideas and approving the words — and hand the typing, reformatting, and posting to the pipeline. Build it once, and "what do I post today?" becomes a question your system answers before you ask it.