Agent 1

The Researcher

The researcher's job is simple: go find relevant, current information on a topic and compile it into a structured brief that your writing agent can work from.

What to connect

• Web search (required). This is your baseline. The agent searches the internet for recent news, trends, and developments related to your content topics.
• Twitter/X (optional but recommended). If your niche has active conversations happening on X, connecting it gives your researcher real-time pulse on what people are talking about.
• Custom source lists (this is where it gets powerful). Give your researcher a list of specific websites, blogs, newsletters, and social media accounts you want it to check. Think of the creators, publications, and thought leaders in your space that consistently put out good stuff. Your agent will check them every time it runs.

How to configure it

In your researcher's SOUL.md (the file that tells the agent who it is and what to do), include:

• A clear description of your niche and audience
• The specific topics and themes you cover
• Your custom source list (URLs and account handles)
• Instructions on output format: you want a structured research brief, not a wall of text. The brief should include the topic, key findings, relevant stats or quotes, source links, and a one-line summary of why this matters to your audience.

Output: A clean research brief that gets passed directly to your writing agent. Think of it like handing a research packet to a human writer. The better the brief, the better the draft.

Agent 2

The Writer

The writer takes the research brief and turns it into a finished content draft in your voice. This is where most people get lazy and end up with generic AI slop. Don't be most people.

Brand voice document

This is the most important file in your entire system. Write a document that describes exactly how you sound. Include:

• Your tone (casual? professional? somewhere in between?)
• How you structure ideas (short paragraphs? story-driven? list-based?)
• Words and phrases you actually use
• Who your audience is and how you talk to them
• Your perspective and beliefs on your topic area

The more specific this document is, the closer the output will sound to you. "Write in a friendly tone" is useless. "Write like a smart friend explaining something at a coffee shop, not a LinkedIn thought leader performing authority" is useful.

Example content

Give the agent 5-10 pieces of content that represent your best work or the style you want to hit. These can be your own posts, or content from creators you want to emulate. The agent uses these as calibration for voice, structure, and quality.

The do-not list

This is your quality filter. List every word, phrase, and pattern you never want to see in your content. Some starters:

• No em dashes (the single biggest AI writing tell)
• No words like "delve," "landscape," "leverage," "foster," "robust," "seamless"
• No "in today's [anything]" or "at the end of the day"
• No "moreover," "furthermore," "additionally"
• No stacked single-sentence paragraphs
• No generic transitions

Build this list over time. Every time you see your agent produce something that makes you cringe, add the offending pattern to the list.

Platform optimization

Tell the writer which platform this content is for. A LinkedIn post has different structure, length, and conventions than a tweet or an email newsletter. You can run the same research through multiple platform-specific writers if you want content for more than one channel.

Auto-Generated Graphics

If you want your agent to produce visuals alongside the written content, add a Remotion or Nanobanana configuration with your brand kit.

After your writer produces a draft, the system generates branded graphics (post images, carousel covers, quote cards) that match your visual identity automatically.

What you need

• Your brand kit (colors, fonts, logo files)
• A Remotion or Nanobanana template configured with your brand assets
• Instructions in the writer's SOUL.md to trigger graphic generation after drafting

This is optional but it turns your content pipeline from "drafts waiting for you" into "ready-to-post packages waiting for you."

Agent 3

The Reviewer

The reviewer is your quality gate. It reads every draft before it goes into your pipeline and checks it against your standards. Think of it as a copy editor who knows your voice, your rules, and your pet peeves.

What it needs access to

• Your brand voice document (same one the writer uses)
• Your do-not list (this is the big one)
• Your example content (so it can compare the draft against your actual voice)

What it checks

• Does the draft sound like you, or does it sound like AI?
• Are there any words or phrases from the do-not list?
• Is the structure right for the target platform?
• Is the content specific and useful, or vague and generic?
• Are there any factual claims that seem off or unverified?

Output: Either a clean, approved draft that moves into your Notion pipeline, or a flagged draft with specific notes on what needs fixing. You can configure this as a hard gate (nothing passes without approval) or a soft gate (flags issues but still moves the draft through with notes).

The reviewer doesn't need to be creative. It needs to be picky. That's why it works well as a separate agent with a single, focused job.

Model recommendation

Claude Sonnet Across All Three

Run all three agents on Claude Sonnet. Here's why.

Sonnet hits the sweet spot of quality, speed, and cost for content work. It's smart enough to synthesize research, match your voice, and catch subtle quality issues. It's fast enough to run a full pipeline in minutes. And it's cheap enough that running it nightly won't drain your API budget.

You might be wondering about Opus (the most capable Claude model). Opus is better for deep reasoning and complex problem-solving, but for a nightly content pipeline, Sonnet gets you 95% of the quality at a fraction of the cost and latency. Over weeks of daily runs, that difference adds up.

If your content is highly technical or your voice is really hard to nail, you could upgrade the writer to Opus. But start with Sonnet across the board and only upgrade if you're seeing quality gaps.

Delivery

Scheduling and Notion

Setting your schedule

OpenClaw runs as a background daemon, so you can set your content pipeline to run at any frequency. Daily, every other day, weekly, whatever fits your posting cadence. Configure the cron schedule in your OpenClaw settings and the system runs automatically.

Connecting to Notion

This is what makes the whole thing feel like magic. Your agent connects to Notion via MCP (Model Context Protocol), which gives it structured access to your databases. When the pipeline finishes, the approved drafts land directly in your Notion content calendar with the right properties filled in (status, topic, platform, date).

You wake up, open Notion, and your content pipeline has new drafts waiting for you. Review them, tweak anything that needs a human touch, and post. That's it.

The build

Claude Code Setup Prompt

Copy the prompt below and paste it into Claude Code. It will walk you through building the entire 3-agent system interactively, asking you questions about your brand, your voice, your sources, and your preferences along the way.

If you don't have OpenClaw set up yet, start with the OpenClaw Secure Setup guide first.