Semrush has hundreds of weblog posts, and a variety of them are informational items readers depend on to study subjects associated to search engine optimisation, AI visibility, and content material. Maintaining these articles present and on the high quality bar Semrush is thought for is a big and ongoing job.
For some time, I attempted to resolve sustaining our informational content material with an n8n workflow. It labored for analysis however broke at drafting.
So, I rebuilt the pipeline in Claude Code. This one handles each analysis and drafting.
Here is why I made the decision to change from n8n to Claud Code, how the brand new system works, and what modified for our workforce.
What stored breaking with n8n
Updating an present article is 2 jobs in a single: an audit and a surgical rewrite.
You need to determine what’s stale, the place rivals have moved, what the AI search panorama now expects, which new product capabilities to weave in, and methods to replace the piece with out touching what’s nonetheless working. Multiplying that by a backlog within the a whole bunch means the workflow needs to be quick, correct, and constant.
My first try and streamline this work was an n8n workflow.
The analysis half labored. For every article, it pulled collectively:
- Complete SERP knowledge for the key phrase
- The highest-ranking competitor articles
- An embedded area intelligence (EDI) scan evaluating our article in opposition to these rivals
- Google’s AI Overview for the question
- Associated searches Google surfaces
- Inner linking alternatives throughout our personal content material
However the drafting by no means labored.
The drafts got here again considerably near what I used to be in search of, however by no means shut sufficient to publish.
The voice was off. The construction ignored the fashion information. The language was fluffy and verbose. And worst of all, there have been hallucinations — the AI generally described Semrush options that do not exist, and in convincing element.
I attempted every thing I may consider to enhance the output. Utilizing totally different AI fashions. Tightening the prompts. Splitting drafting into smaller steps. Giving it the fashion information. Giving it extra previous drafts as examples.
None of it produced constant, high-quality outputs. I might get a suitable draft as soon as, then the following run could be mistaken in a brand new manner.
Finally I ended making an attempt to repair the content material I used to be getting from n8n. The analysis half nonetheless gave us info for briefs the workforce may write from, so we stored that working and set the drafting apart.
However I couldn’t cease serious about why the drafting stored failing.
It seems the failure was structural all alongside. n8n is nice at chaining API calls — fetch this, remodel that, and ship it onward.
Drafting an article, nevertheless, requires editorial reasoning — judgment calls about voice, construction, and what to alter. That type of reasoning wants to think about the entire article directly, plus reference materials just like the fashion information and previous examples obtainable as selections get made.
Workflow instruments merely aren’t constructed for that.
Why I switched to Claude Code
I wanted one thing that would do actual editorial work, like learn the unique article, perceive the intent behind the question, and make calls about what to alter and what to go away alone.
I checked out a number of choices and stored coming again to Claude Code.
Here is what made it match:
Claude Code is an agent that runs inside a folder in your pc. The pipeline is that folder. The fashion information, previous drafts, the analysis output, and the article being up to date are all information inside it.
Claude Code reads what it wants when it wants it, and the work it does turns into one other file the following step can use.
The structural distinction from n8n is in how the AI matches into the workflow. In n8n, you construct the workflow upfront, and the AI does one particular step, like writing a piece or summarizing knowledge.
In Claude Code, the AI runs the workflow itself, studying the information, deciding what to do, and writing the outputs. Mixed with ability directions that inform it what to do at every step, Claude Code has each the context drafting wants and the constraints that preserve it from going off the rails.
That is what made the distinction.
The AI had entry to what it wanted when it wanted it, and an outlined job at every step. The work it produced was a file the following ability may decide up and a author may open later to verify.
I rebuilt the entire pipeline in Claude Code, together with the API calls that had been working high-quality in n8n. With every thing in a single folder, the drafting step may learn the analysis output, the unique article, previous drafts, and the fashion information at any time when it wanted them.
And it labored.
The pipeline produces drafts our writers can edit and publish, and a path of information they’ll verify when one thing seems to be off.
9 abilities, finish to finish
The pipeline I in-built Claude Code is 9 abilities, chained collectively by a grasp script that runs them so as.
I give it the URL of the article I wish to replace and a goal key phrase, and I get again a draft. The draft goes by means of our regular editorial workflow the identical as every other article: assessment, revisions, modifying, and pictures. Our workforce makes each editorial name.
Listed here are the 9 abilities:
- Fetch the dwell article
- Analysis the SERP and rivals
- Run an EDI semantic similarity verify in opposition to our present piece
- Synthesize an replace plan
- Determine outdated content material
- Audit product mentions
- Draft the updates
- Generate a side-by-side comparability of the unique and the brand new draft, with adjustments highlighted
- Format the consequence for publishing
I stored it at 9 abilities on objective. It was the smallest quantity that gave me a definite ability for each choice the pipeline wanted to make.
And one design selection turned out to be actually necessary. Each ability saves its work to a file earlier than the following one runs.
These information are what I name the pipeline’s artifacts. They embrace the analysis, the plan, the draft, and the side-by-side comparability. Saving every step as a file means any single ability might be re-run with out beginning over, and anybody can open the information to verify when a draft seems to be off.
What modified when the Claude Code pipeline ran
Two issues modified when the Claude Code pipeline began working:
- The hallucinations the AI nonetheless often produced grew to become straightforward to catch
- The drafts began studying like we wrote them
Any AI era step can hallucinate generally. The pipeline is constructed to catch them quick.
Dana — one in all our contributors — was reviewing a draft and bumped into plausible-looking directions for a function that does not exist. The type of error that, within the previous n8n model, would have both slipped by means of or value twenty minutes of cross-checking.
She opened the side-by-side diff, appeared on the similar part within the unique article, noticed the unique did not point out the workflow, and changed the fabrication. The entire thing took a couple of minute.
Right here’s what the diff artifact seems to be like:
That is what the artifacts are for. The AI remains to be going to make errors. The pipeline is constructed so a reviewer can catch them and verify in a single minute as a substitute of 20 minutes.
The larger story is what occurred throughout runs.
For months, I might been making an attempt to get the drafting step to provide one thing that learn like Semrush. Which means the fitting strategy to voice, tone, construction, and the way we describe our personal merchandise. In n8n, I might get a draft that perhaps nailed a type of issues and missed three others. And the following run, I’d get a special mixture.
However in Claude Code, three runs with small changes between them bought me there. By the third, the drafts had been constantly robust.
The voice matched the present article. The construction adopted our fashion information. The tone was Semrush. The model positioning was proper. The AI bought the product descriptions appropriate. The identical type of errors did not preserve displaying up in other places.
This was the half I hadn’t anticipated. Months of changes in n8n hadn’t gotten me right here. Three runs in Claude Code did.
Dana nonetheless caught issues, however they had been the smaller editorial fixes any draft wants, like sharpening a gap, reframing a piece, or smoothing a clunky transition. The drafts now not arrived with the larger issues n8n had given us, just like the mistaken voice, ignoring the fashion information, or fabricated Semrush options.
Dana’s suggestions after a number of runs was that the writing was a lot better than what we might produced earlier than. And the side-by-side view was truly helpful.
What ended up mattering
Three issues held up throughout each run.
- Drafting wants full context. Treating the LLM as one step in a workflow provides you inconsistent writing. The drafting work has to see the article, the fashion information, and the analysis on the similar time.
- The path of information is the system. Each ability saves its work earlier than the following one runs. That path is how our workforce catches issues, and the way I can re-run any single step with out beginning over.
- Fewer abilities, extra refinement. 9 coated the work. Each time I have been tempted so as to add a tenth ability, the fitting transfer has been to sharpen one of many present 9.
The pipeline is working, the workforce is utilizing it, contributors are saving substantial time, and the suggestions has been extra optimistic than something we have had with AI-generated content material.
In case you’re hitting a high quality ceiling with AI content material, begin by asking the place your AI is making its writing selections. In the event that they occur inside a workflow step, that is the place the ceiling is coming from.
Transfer the drafting work someplace the AI can learn your information instantly. That is likely to be an agent like Claude Code or any instrument that provides the AI persistent entry to reference materials. That is the transfer that broke by means of the ceiling for us.
