How to build an AI news-curation agent for social media
This guide is based on the How AI Maturity Levels Transform Your Marketing Work Week webinar, published on the AI Lab by ActiveCampaign.
Get the quick-start guide
What will I accomplish with this guide? By the end of this guide, you’ll have a two-agent workflow that scans industry news every morning, judges relevance for your audience, and drafts platform-specific LinkedIn and Twitter posts in your brand voice, with you reviewing the output before it ships. It’s based on the system Maddy Osman built for The Blogsmith, the same workflow that replaced one to two hours of manual scanning every morning.
Before you start, you’ll need:
- A Zapier account on a plan with Agents access (or an n8n instance — Maddy ships a version of this in both tools)
- A Google News API integration or equivalent search connector available in your automation tool
- An Airtable base (or any structured store) where each news item can live as a record
- A written brand style guide and a short “quick tips” version (avoid jargon, encourage engagement, etc.)
- A Slack workspace, an email inbox, or a social scheduler where the drafts will land for review
Quick reference
- Total time: 2–4 hours to build the Zapier version; longer for n8n if you want stricter formatting control
- Tools needed: Zapier Agents (or n8n), Google News search API, Airtable, Slack/email/social scheduler, your brand style guide
- Key output: Daily Slack or email digest with LinkedIn and Twitter drafts written in your brand voice, ready for human approval
Watch this section
For full context on the following topics, watch these sections of the webinar:
- The two-agent architecture and why Airtable sits in the middle — [19:25]–[21:39]
- Live walkthrough of agent one in Zapier (search → analyze → store) — [22:25]–[24:30]
- The n8n rebuild and why determinism matters — [25:30]–[28:00]
The workflow
Phase 1: Define the editorial filter before you build anything
After this phase, you’ll have: a written set of relevance rules and a brand-voice document the agents can reference.
- Write your audience in one sentence: Maddy’s is “CMOs at growing B2B brands.” Specificity here is what makes the relevance check work.
- List the keywords you want to monitor: Maddy started with “AI workflow automation.” Pick three to five.
- Draft the relevance questions the agent will ask each story: Who benefits? Why should a CMO care? Does it align with our expertise?
- Save your full style guide as a knowledge document: the long version goes into the agent as reference material.
- Write a short “quick tips” version of your style guide: five to seven bullets that hit the highlights (avoid jargon, end with a thought-provoking question, etc.). Maddy learned the long doc alone wasn’t enough; the agent needs a TL;DR it can hold in working memory.
Phase 2: Build agent one — find and analyze news
After this phase, you’ll have: a Zapier Agent that runs every morning, finds relevant stories, and saves each one to Airtable as a structured record.
- Create a new Zapier Agent and set the trigger to run daily at 8 a.m.
- Connect Google News search via Zapier’s API connector: wire in your keyword list from Phase 1.
- Add Zapier’s AI connector as the analysis step: instruct it in natural language to score each article against your relevance questions.
- Tell the agent what to extract: a one-paragraph summary, why a CMO should care, the source URL, and the publish date.
- Pipe the output to Airtable: create a “Suggested posts” view where every relevant story becomes a new record. The Airtable layer is what stops you from re-sharing the same story tomorrow.
Sometimes it’s nice to have a storage step before you have the execution step, depending on what you’re building.
Phase 3: Build agent two — draft platform-specific posts
After this phase, you’ll have: a second agent that picks up new Airtable records and drafts a LinkedIn post and a Twitter post for each.
- Create a second Zapier Agent triggered on new Airtable records.
- Attach your full style guide as a knowledge file and your “quick tips” as inline instructions.
- Tell the agent the platform constraints: LinkedIn can run longer with line breaks; Twitter caps at 280 characters and rewards punchier framing.
- Ask for both posts plus the rationale: Maddy specifically asked the agent to include why it chose each story so she could approve faster.
- Send the output to Slack: include the source link, the LinkedIn draft, the Twitter draft, and the rationale in one message.
Phase 4: Add a human-in-the-loop review step
After this phase, you’ll have: a daily approval ritual that takes minutes instead of an hour.
- Block 10 minutes on your morning calendar to review the Slack digest.
- Approve, edit, or skip each post: Maddy keeps the AI’s first draft as a starting point and edits in her voice, the same pattern Amanda uses for sales follow-ups in the companion guide.
- Track skip reasons for one week: if you keep killing the same kind of story, tighten the relevance prompt in agent one.
Phase 5: Optional — rebuild in n8n for stricter control
After this phase, you’ll have: a more deterministic version of the same workflow, with the option to push approved drafts straight to a social scheduler or to email.
- Mirror the Zapier flow in n8n: same trigger, same search, same Airtable storage, same drafting step.
- Replace natural-language instructions with explicit nodes where output formatting matters: Maddy moved to n8n because Zapier’s natural-language steps could vary day-to-day; n8n let her lock the output shape.
- Add a delivery branch for each stakeholder: Maddy ships her own version to Slack and a client version to email, with an optional draft straight into a social scheduler.
- Keep human approval in the loop: even with the scheduler integration, every draft sits as a draft until someone clicks publish.
AI does the heavy lifting, but I have the final approval. You want to trust the output, but you also want to maintain quality control before it becomes client or public.
Related
More data from the AI Lab.