Anastasia Braitsik is a leading figure in the digital marketing landscape, currently serving as an expert in SEO, content marketing, and data analytics at Semrush. With a career dedicated to bridge the gap between technical data and creative storytelling, she has pioneered methods to maintain high-quality informational content at a global scale. Her work often involves navigating the complex intersection of automation and editorial integrity, ensuring that as technology evolves, the brand’s voice remains authentic and reliable. Recently, she has been at the forefront of shifting from traditional workflow automation to more sophisticated, agentic AI systems to solve the persistent challenges of content aging and accuracy.
In this discussion, we explore the fundamental limitations of traditional automation tools when applied to nuanced creative tasks and why standard “chaining” of API calls often fails to capture a brand’s unique persona. We delve into the tactical shift from linear workflows to agent-based systems that utilize “editorial reasoning” to handle large backlogs of content updates. The conversation highlights the importance of persistent context, the strategic design of a nine-skill pipeline, and the creation of “artifacts” that allow human editors to verify AI outputs in seconds. By examining the transition from n8n to Claude Code, we uncover how a folder-based intelligence system can eliminate hallucinations and produce drafts that truly mirror the quality of professional human writers.
Many automated workflows fail when moving from data retrieval to creative drafting because they treat writing as a series of linear steps. From your experience rebuilding the content pipeline at Semrush, why did the initial n8n setup struggle to produce publishable content even when the research phase was successful?
The failure of the initial n8n workflow was deeply rooted in its structural design, which simply wasn’t built for the complexities of a “surgical rewrite.” Updating an existing article requires two distinct jobs: a comprehensive audit of what is currently stale and a creative integration of new product capabilities without damaging the elements that are already working. While n8n was excellent at pulling together data points—such as SERP information for keywords, top competitor articles, and Google’s AI Overviews—it treated the actual drafting as just another isolated step in a chain. Because it functioned by passing data from one node to the next without a holistic view, the resulting drafts were often fluffy, verbose, and completely ignored our internal style guide. Most concerningly, the system would hallucinate, describing Semrush features that don’t even exist in very convincing detail, because it lacked the “editorial reasoning” to cross-reference the research with our brand reality. We found that even when we tightened prompts or split the drafting into smaller steps, the output remained inconsistent; we might get one acceptable draft, but the next would be wrong in an entirely new and unpredictable way.
You eventually transitioned to Claude Code to manage the drafting process. How does this agentic approach change the way the AI interacts with your brand’s style guide and existing content compared to a standard workflow tool?
The shift to Claude Code represented a fundamental change in how the AI “lives” within our workflow, moving it from a single step in a process to the entity running the process itself. Claude Code operates as an agent within a specific folder on a computer, and that folder essentially acts as its comprehensive memory bank. Inside this directory, we place the style guide, past successful drafts, the raw research output, and the original article that needs updating, allowing the AI to read what it needs exactly when it needs it. In a tool like n8n, you build the logic in advance and the AI performs a specific, narrow task, but in this new system, the AI makes judgment calls about voice and structure based on all those files simultaneously. This persistent access to context means the AI isn’t just following a prompt; it is using “editorial reasoning” to decide what to change and what to leave alone. This structural difference is what finally allowed us to break through the quality ceiling, resulting in drafts that finally matched our brand positioning and described our products with 100% accuracy.
The pipeline you built consists of nine distinct skills. Could you walk us through how these stages interact and why you decided to have the system save “artifacts” at every single step?
I designed the pipeline around nine specific skills to ensure every editorial decision had its own dedicated moment for refinement and verification. The process begins with fetching the live article and researching the current SERP and competitors, followed by a semantic similarity check to see how our piece stacks up against the current landscape. From there, the system synthesizes an update plan, identifies outdated content, and audits product mentions before finally drafting the updates and generating a side-by-side comparison. A critical design choice was requiring every single skill to save its work to a file—an “artifact”—before the next skill begins its run. These artifacts, which include the research notes, the specific update plan, and the final draft, create a transparent trail that our human editors can inspect at any time. This means if a draft looks slightly off, a writer can open the specific research file or the update plan to see exactly where the logic shifted, and I can re-run any individual skill without having to restart the entire nine-step process from scratch.
One of the most interesting parts of your new system is the “side-by-side” comparison. How has this specific feature changed the daily life of your editorial team and their ability to trust AI-generated drafts?
The side-by-side comparison, or the “diff artifact,” has completely transformed the way our contributors interact with AI, turning a high-friction review process into a streamlined editorial task. Previously, checking an AI draft for hallucinations or style errors could take twenty minutes of tedious cross-referencing against the original text and the product database. Now, editors like Dana can open the comparison file and immediately see exactly what was changed, added, or removed in a clear, highlighted format. For instance, if the AI fabricates a set of instructions for a non-existent feature, the editor can see at a glance that the original version didn’t contain that information and correct the fabrication in about sixty seconds. This system acknowledges that AI will still make mistakes, but it builds a safety net that makes those mistakes incredibly easy to catch. The team now feels empowered because the drafts no longer arrive with massive structural problems or “off-voice” tones, allowing them to focus on high-level editorial improvements like sharpening transitions or reframing a section’s opening.
What is your forecast for the future of AI-driven content maintenance?
My forecast is that we are quickly moving toward a world where the most successful content teams will stop using AI as a simple “content generator” and start using it as an integrated “contextual agent.” We will see a shift away from rigid, linear workflows that treat writing as a data-transformation task and move toward environments where the AI has persistent, deep access to a brand’s entire knowledge base. This will lead to a significant reduction in the manual labor required for content audits, as these agentic systems will be able to autonomously monitor the SERP landscape and suggest surgical updates in real-time. I believe the “hallucination problem” will largely be solved not just by better models, but by better systems architecture that prioritizes “artifacts” and human-in-the-loop verification. Eventually, the distinction between “AI-written” and “human-written” will fade, replaced by a standard of “expert-verified” content where the heavy lifting of research and drafting is handled by agents, but every strategic and editorial call remains firmly in human hands.
