Neos Conf 2025 Talk by Felix and Roland

Letting Editors Go Wild: 
An Editor-Centric Approach to AI Agents

I know some of you might be thinking, "Oh no, another AI talk." I get it. The hype is exhausting. We've all sat through presentations promising AI will revolutionize everything, only to return to our desks and find the same old workflows waiting for us.

But here's the thing: looking past the hype, AI is genuinely useful. I'm not saying AI is living up to the breathless predictions of tech evangelists. What I am saying is that there's real utility in AI, independent of the hype cycle. So let's figure out what this utility actually is and how we can harness it in ways that matter.

The Humbling Art of Time Capsules

Let me share something that keeps me grounded: time capsules. The concept is simple—you write down predictions for what will happen in six months or a year, seal them away, and check later to see how wrong you were. It's a humbling exercise that prevents us from constantly rewriting history in our minds, convincing ourselves we "saw it coming all along."

My track record? Consistently terrible.

In 2024, after building a voice-based AI coach with semi-decent results, I confidently predicted real-time voice interaction was at least a year away. Less than three months later, OpenAI released ChatGPT's advanced voice mode, making my prediction look antiquated.

In November 2024, I bought into the slowing open of new training ideas and narrative that generated data will worse than real data. Then December rolled around, and OpenAI dropped o1-pro. Reluctantly, I tested it despite the $200 monthly price tag and discovered an entirely new level of reliability and deep reflection from first principles that unlocked capabilities I hadn't thought possible.

But perhaps my most significant misprediction involved AI coding. For months, I argued against the hype around AI-assisted development. Then, this February, I forced myself to spend a weekend building a project without writing any code—just precisely instructing AI what to do. The first hour felt bizarre, almost wrong. But by day's end, I understood where this is all heading.

The kicker? Last week, I heard Adam Wathan, creator of Tailwind CSS, describe an almost identical journey. We're not alone in this revelation.

At this state, Engineering beats Intelligence

Here's what I've learned from these mispredictions: it's extremely sad to still hear people dismiss AI as "stochastic parrots" when these systems are clearly capable programmers. There's so much we can build with this technology, and I don't understand developers who aren't using it daily.

But here's the crucial insight I missed initially: if you compare tools like Cursor or Windsurf from six months ago to now, the biggest quality improvements aren't coming from better language models. They're coming from brilliant engineering. The LLMs are already good enough to be useful—what matters now is how we build around them.

Think about it: there are billion-dollar companies building AI coding agents. Companies like Cursor and Windsurf are reshaping how developers work. But in the CMS space? We haven't seen similar investment yet.

It's coming.

The question isn't whether a CMS will become AI-first. The question is which CMS will become the Cursor of content management.

How to not build AI products

Before we dive into solutions, let's examine how not to build AI features. Gmail's "Help me write" feature is a masterclass in missing the point.

Picture this: Google's flagship AI model, Gemini, offers to help you write emails from scratch. Sounds helpful, right? Here's what actually happens: you type a prompt explaining what you want the email to say. Gemini produces a perfectly reasonable draft that sounds like it was written by a corporate communications robot who's never met a human being.

The tone is generic. The language is stiff. And here's the real issue: writing the prompt often takes longer than just writing the email yourself. As Pete Koomen brilliantly put it, "Remarkably, the Gmail team has shipped a product that perfectly captures the experience of managing an underperforming employee."

Why does this happen? Because Gmail's approach treats the AI like a black box. Users have no control over the system prompt that shapes how the AI writes. There's no way to teach it your voice, your style, your brand. It's the equivalent of hiring a ghostwriter who refuses to listen to how you actually communicate.

The likely system prompt probably looks something like: "You are a helpful email-writing assistant responsible for writing emails on behalf of a Gmail user. Follow the user's instructions and use a formal, businessy tone and correct punctuation so that it's obvious the user is smart and serious."

This completely removes user control. The result? Generic AI slop. And that's the opposite of everything we believe in when building content management systems.

Don't Be Gmail

We need to build integrations where editors maintain full editorial control. Good AI UX design means editable prompts for end users. When an AI agent acts on my behalf, I should be able to teach it how.

For a CMS like Neos, this means the editor must be in full control of the system prompt. They define the brand, the tone, the rules. They teach the AI to be a true co-pilot for their brand.

This isn't about AI replacing editors. It's about editors becoming prompt architects, allowing them to scale their expertise across their entire content operation. What if YOU could teach the AI your brand voice, your style, your rules?

Beyond the Marketing Terms

Yes, "AI Agents" is a marketing term. Let's move past the definitional debates and talk about concrete things we can build. What matters is utility: tools that can read, understand, and transform content according to YOUR rules.

Think of them as extremely capable interns who actually read the instructions you give them—a rarity in any universe.

The Pressure on Modern Editors

Content creators today face an impossible challenge. They're expected to produce more and more content, competing for attention with Elon Musk's latest tweet and an infinite scroll of viral videos. Staying relevant feels like running up a down escalator.

Traditional AI tools aren't helping. They produce generic outputs that feel artificial and disconnected from brand identity. But there's another way.

Two Paths to Valuable Content in the AI Era

When we think about what makes content valuable, two paths emerge:

  1. Depth & Excellence: Like reading a book from a legendary software architect—rare, profound, transformative.
  2. Extreme Personalization: Like getting feedback from a skilled architect who has read ALL your code and understands your specific context.

AI enables this second path at scale. It can tailor content precisely to each reader's context, making even moderately good writing valuable if it answers YOUR specific question in exactly your preferred style. This is where tools like Deep Research are heading—superior tailoring that transforms generic content into personalized insight.

But here's the opportunity: most people will need years to adopt these AI-powered reading tools. This creates a window where editors who embrace AI-assisted personalization can provide extraordinary value to their audiences.

Our Journey

For the past two years, my colleagues Daniel, Felix, and I have been exclusively focused on the intersection of AI and content management. We've built production systems for clients like Wiener Zeitung and Puls4, and we're here to share the most important patterns we've discovered.

More importantly, we've verifiably saved thousands of hours for editors and quality ensuring. Not through magic, but through reliable AI-enhanced workflows.

Building a Brand Guard

Enough theory. Let's build a Brand Guard—a practical example of an editor-centric AI agent.

Every organization has brand guidelines. Think about something as simple as how you write your company name. Is it "Neosidekick" or "NEOSSidekick" or "NEOSidekick"? Now multiply that by hundreds of rules about tone, terminology, and style. Checking all of this manually is tedious and error-prone.

What if every content change automatically triggered a brand check, with an AI agent suggesting improvements without interrupting the editor's flow?

Let me walk you through how we actually built this. The concept is straightforward:

  1. Your editors change content, which triggers a webhook
  2. A workflow tool receives this webhook and passes it to an AI agent that knows your brand guide
  3. The agent checks the writing style and suggests improvements if necessary
  4. The editor is notified and can decide whether to accept or reject the recommendations

Why N8N?

For the orchestration layer, we chose N8N. Why? First, it's open source—a value we all share in the Neos community. You can run it in their cloud, but you can also host it on your own servers. The workflow files can be downloaded as JSON files, keeping you in full ownership of your automation logic.

You can connect almost everything—OpenAI, your CRM, analytics tools, or Neos itself. And here's what I love: it uses visual workflows that non-developers can understand and maintain, while still giving developers the power to drop into code when needed.

Copy and adopt the Brand Guard for your own needs

Neos AI package and N8N Workflow Template

The agent creates a new workspace—not touching the live content—and suggests the correction. I receive a Slack notification with a direct link to review the changes. One click takes me to Neos's workspace module where I can accept or discard the suggestion.

But let's push this further. Microsoft's style guide contains over 1,100 pages. No human can remember all of that while writing. In our presentation we show, how OpenAI o3 can handle this hugh document and catches subtleties that would slip past most editors.

The Technical Implementation

The magic happens through a few key components:

  1. Neos Package: Triggers webhooks on content changes and provides endpoints for workspace manipulation
  2. N8N Workflow: Receives the webhook, formats the data, and coordinates the AI agent
  3. ChatGPT Integration: Evaluates content with a carefully crafted system prompt
  4. Workspace API: Creates isolated environments for proposed changes
  5. Notification System: Alerts editors through Slack, email, or other channels

The beauty is that editors never lose control. Every suggestion goes through human review. It's like having a skilled intern who makes recommendations but never publishes without approval.

Building a Junior Editor

Brand checking is just the beginning. What if we could give editors an AI assistant that helps with more complex tasks? Enter the Junior Editor—an agent that can traverse your content structure, make bulk changes, and handle migrations that would typically require developer intervention. Instead of a fixed workflow, editors interact through a simple chat interface. Watch this:

Download the N8N Junior Editor Agent demo

Turning This Into Production Code

Before you rush to implement this, let's talk about doing it safely:

  1. Security First
    The agent must never have more permissions than the editor using it. This prevents prompt injection attacks from escalating privileges. If an editor can't publish directly to production, neither can their AI agent.
  2. Permissions
    The agent is never allowed to have more permissions than the editor.
    Slack approvals only when you have small trustworthy teams.
  3. Content PRs
    Think of AI suggestions like pull requests. It's the editor's responsibility to review and approve changes. Just as developers learned to work with code PRs, editors must learn to review content PRs. If you blindly accept suggestions with errors, that's on you—not the AI.
  4. Bonus: Build as MCP Tools
    By implementing your tools using the Model Context Protocol, they become reusable across different AI systems and workflows.

Writing Perfect Prompts

System prompts are where the magic happens. Here's what we've learned:

  • Identity and Context First
    Start by defining the AI's identity, current date, and key facts. This reduces confusion and improves consistency.
  • Be Specific
    Don't say "use friendly tone." Say "address readers as 'you', avoid third person, use contractions like 'you're' instead of 'you are'."
  • Show Examples
    Include 3-5 examples of good brand writing. The AI learns patterns better from examples than from rules.
  • Use Declarative Rules
    "Always talk in direct second-person" is clearer than "try to be conversational."
  • Implement Triggers
    Use explicit if-then blocks for edge cases. "If discussing pricing, always include 'excluding VAT' after amounts."

Add your software engineering skills:

  • Regular Evals
    Track how often suggested rewrites are accepted or rejected. If acceptance is low, adjust rules accordingly.
  • Combine models
    AI Tools are like lego blocks, you can build more interesting things with different colors. (Andrew Ng)

Solutions You Could Build

The possibilities are endless. Here are proven patterns that I'm confident can be built sucessfully:

Brand Guard: Automatically checks tone, terminology, and style consistency. 

SEO Junior: Adds missing meta descriptions, suggests internal links, optimizes heading structures.

Accessibility Adviser: Ensures alt text exists, and validates heading hierarchy.

Engagement Improver: Weekly analysis of analytics data, suggests A/B tests for underperforming content.

Trend Spotter: Daily scan of news and social media, drafts timely content ideas for editorial review.

Translation Maintainer: Keeps multilingual content synchronized while preserving brand voice across languages.

What You Need to Build This

An AI agent is simply an LLM, state management, and tools. Here's your shopping list:

  1. LLM & Guidance
    Understand what model to pick, if you need orchestration, how to write good system prompts, what to keep flexible for editor.
  2. Memory Management & State Architecture
    Design strategies for working memory, episodic memory, and long-term memory systems, plus security considerations for memory poisoning attacks and data protection in production AI agent deployments.
  3. Integration Architecture
    Connect your CMS, analytics, CRM, and other tools. Rate limiting and error handling become critical at scale.
  4. Production AI Agent Security
    Security architecture covering prompt injection defense, data exfiltration prevention, compliance and AI-specific threat models

The Path Forward

We're at an inflection point. Just as Cursor and Windsurf transformed coding, AI-powered CMS tools will transform content creation. The question isn't if—it's who will build the defining tools of this new era.

While widespread adoption will take years, early movers can establish significant advantages. Small automations compound into massive time savings. A 5-minute daily task automated saves 20 hours per year. Multiply that across your editorial team and dozens of workflows.

Start small. Pick one tedious task. Build an agent for it. Measure the time saved. Use that success to fund the next automation.

Your Next Steps

Ready to let your editors go wild? The tools exist. The patterns are proven. The only question is: what will you build first?

Connect with the community: Join Slack #ai and share what you want to build

Start experimenting: Download our Neos AI package, N8N Brand Guard Workflow, N8N Junior Editor Agent and adapt it to your needs.

Think bigger: If you're interested in building MCP tools for Neos, ping @rolandschuetz on Slack.


The ideas here are mostly a collection from other great minds, condensed to what I find most important. And since I'm not the best writer, I used a lot of AI to formulate it. I hope it's of value to you.

Remember, this isn't about AI replacing editors. It's about giving editors superpowers. It's about turning repetitive tasks into creative opportunities. Let's build cool stuff!

PS: We are working on a production-ready Brand Guard and SEO assistant for NEOSidekick. Hopefully we can share more learnings soon.

Resources:

Background materials: