The AI Content Agent: Build a Content Business That Runs Itself

1. What you are actually selling

Every local business owner you will ever pitch has the same problem: the raw material for good content already exists — customer questions in the inbox, photos on a phone, a new product line, a seasonal offer — but turning it into posts, articles and emails takes hours they do not have. A content agent is the machine that closes that gap: it collects the raw material, drafts in the business's voice, repurposes each piece for every channel, and queues everything for one fast human approval.

This is Guide 3 of our AI agent business series. The sales agent and the receptionist agent convert attention into bookings; the content agent is what manufactures the attention in the first place. One honest boundary before we build: a content agent "runs itself" only inside its guardrails. Research, drafting and scheduling are automated. Publishing is not — a person approves every piece, and that approval gate is precisely what makes the service safe to sell.

2. The pipeline: research → draft → repurpose → schedule

Four stages, each with a defined input and output. If a stage's output is not written down, the next stage hallucinates — so every stage hands over a document, not a vibe.

• Research. Inputs: the client's notes and inbox questions, the approved RSS sources, competitor pages worth answering. Output: a ranked topic list with the source for every claim, scored by customer relevance — not by what is trending globally.
• Draft. The model writes one anchor piece (a blog post or a long-form answer) from a brief: topic, audience, the facts it may use, the offer it may mention. Output: a draft with every factual claim traceable to the research document.
• Repurpose. The approved anchor becomes native pieces per channel: short social posts, a newsletter section, a video script outline. Each is rewritten, not truncated — a caption is not a paragraph with hashtags.
• Schedule. Approved pieces enter the publishing queue with dates and channels. Output: a calendar the owner can see, pause or edit at any time.

{
  "stages": ["research", "draft", "repurpose", "schedule"],
  "sources": {
    "rss": ["industry-news-feed", "supplier-blog"],
    "client_inputs": ["owner-notes-doc", "inbox-questions"],
    "competitors_to_answer": ["rival-blog"]
  },
  "gates": { "publish_requires": "human_approval" },
  "cadence": { "anchor_pieces_per_week": 1 }
}

3. The stack: a model, sources, glue and a queue

Nothing exotic. A frontier LLM does the anchor drafting, where voice and judgment live; a cheaper, faster model handles classification, summaries and repurposing variants, where volume lives. Sources connect through RSS and simple HTTP fetches. The glue is a workflow tool — this series uses n8n for its agent workflows, and a comparison of the main options is in our n8n vs Make vs Zapier guide — because the pipeline is triggers, branches and API calls, exactly what those tools are for.

The piece most builds get wrong is the queue. The approval queue is the product the client interacts with: a shared sheet, a Notion database or a simple board where each draft sits with its sources and three buttons — approve, edit, reject. Publishing then happens through the CMS API (as a scheduled post) and the social scheduler's API, and only for items whose status is "approved". If reviewing takes the owner more than a glance per piece, your packaging — not the model — is the problem.

4. The system prompt (steal this)

This prompt runs the drafting stage. The placeholders are the onboarding work: one interview with the owner fills the voice profile, the source list and the banned claims. Adapt the rules, never delete the hard ones.

You are the content agent for {business_name}, a {business_type}
in {city}. You write in the owner's voice: plain sentences, local
specifics, zero hype adjectives.

INPUTS YOU RECEIVE
- Raw material: owner notes, product updates, customer questions.
- Research: only the approved source list. Nothing else exists.
- The voice profile and the banned-claims list for this business.

YOUR JOB, IN ORDER
1. RESEARCH: rank candidate topics by relevance to our customers
   and fit with our offers. Attach a source to every factual claim.
2. DRAFT: write the anchor piece from the brief I give you, using
   only facts from the inputs.
3. REPURPOSE: rewrite the anchor as 3 social posts, 1 newsletter
   section and 1 short video script — each native to its channel.
4. SCHEDULE: place every piece in the approval queue, with sources.

HARD RULES
- Never invent prices, statistics, testimonials or quotes.
- A fact that is not in the inputs gets marked [VERIFY] in the
  draft and is never silently resolved by you.
- Your final state is always "awaiting approval". You never
  publish, post or send anything yourself.
- Output one JSON object per piece: title, channel, body, sources.

5. A worked week: the pipeline in motion

One client, a home-services business, running the default cadence. Days are when things happen; the owner's total time is the last line.

• Monday — research. The agent pulls the week's RSS items, the owner's two voice notes ("we now offer same-day water-heater installs", "Mrs. Alvarez asked if tankless is worth it") and the top inbox questions. Output: 12 candidate topics, ranked; the tankless question scores first because three customers asked versions of it.
• Tuesday — draft and repurpose. The anchor piece is drafted: "Tankless vs. tank water heaters: an honest cost breakdown for {city} homes", with every number sourced or marked [VERIFY]. Repurposing produces three social posts, the newsletter section and a 45-second script outline.
• Wednesday — the approval sitting. The owner opens the queue with coffee: approves most as-is, edits one line (a price changed), rejects one post as off-brand. Total sitting: about 30 minutes.
• Thursday–Friday — schedule and learn. Approved pieces go out on their slots. On Friday the agent files a one-page note: which topics earned replies and questions, fed back into next Monday's ranking weights.

That loop — not any single clever prompt — is the service. The owner experiences it as one short sitting a week; everything else simply appears.

6. Example pricing: a model with labeled assumptions

Stated plainly, so you can audit it: this is an example model, not market research and not a promise of what you will earn. Assumptions: one local-services client; deliverables are 4 anchor pieces, 12 social posts and 1 newsletter issue per month; you run the pipeline for them.

• Setup, one time — example: $750. Covers the onboarding interview, voice profile, source list, pipeline build and queue setup.
• Monthly care — example: $600/month. Covers running and maintaining the pipeline. Inside it, assume roughly $40–80/month of hosting and API costs and about 2 hours of your own review and fixing time; the rest is your margin — check it against your real costs, not this example.
• The client's side of the math: they compare the fee against the hours a staffer or freelancer would spend producing the same approved output, and against the cost of silence — months with no posts at all.

7. Quality rules: the human gate is the product

Five rules keep this service sellable month after month. Approval before publish, always — no client, however busy, gets the auto-publish switch in version one. Facts come from inputs or carry [VERIFY] — a made-up statistic published under a client's name ends the relationship and possibly worse. Voice is checked against samples — keep three owner-approved pieces as the reference set and draft against them. Sources are curated, not accumulated — a feed that produces two bad drafts gets removed, however famous it is. And the pause button is sacred: one click stops the queue, and the client is told where it is on day one.

8. Common mistakes

• Selling volume instead of approval. "Thirty posts a day" is a liability pitch. Clients renew for a queue they trust, not for a firehose they fear.
• Skipping the voice interview. With no voice profile and no approved samples, the agent produces competent, generic text that sounds like everyone's — and no one's.
• Truncating instead of repurposing. A blog paragraph pasted into a social post reads like exactly what it is. Each channel gets a rewrite, or the engagement data will explain it to you.
• Letting the source list rot. Feeds drift, competitors pivot, and last season's topics keep scoring. Review the research output monthly and prune without sentimentality.

Previous in the series: Guide 2 — the receptionist agent. Next in the series: Guide 4 — the trading agent, without the hype. Or return to the series hub for the full roadmap.