I rebuilt our content update pipeline in Claude Code. Here's why.

Semrush has thousands of blog posts, and a batch of them are informational pieces readers trust connected to study astir topics related to SEO, AI visibility, and content. Keeping those articles existent and astatine the value barroom Semrush is known for is simply a important and ongoing job.

For a while, I tried to lick maintaining our informational contented pinch an n8n workflow. It worked for investigation but collapsed astatine drafting.

So, I rebuilt the pipeline successful Claude Code. This 1 handles some investigation and drafting.

Here's why I made the telephone to move from n8n to Claud Code, really the caller strategy works, and what changed for our team.

What kept breaking pinch n8n

Updating an existing article is 2 jobs successful one: an audit and a surgical rewrite. 

You person to fig retired what's stale, wherever competitors person moved, what the AI hunt scenery now expects, which caller merchandise capabilities to weave in, and really to update the portion without rubbing what's still working. Multiplying that by a backlog successful the hundreds intends the workflow has to beryllium fast, accurate, and consistent.

My first effort to streamline this activity was an n8n workflow.

The investigation half worked. For each article, it pulled together:

• Comprehensive SERP information for the keyword
• The top-ranking competitor articles
• An embedded domain intelligence (EDI) scan comparing our article against those competitors
• Google's AI Overview for the query
• Related searches Google surfaces
• Internal linking opportunities crossed our ain content

But the drafting ne'er worked.

The drafts came backmost somewhat adjacent to what I was looking for, but ne'er adjacent capable to publish.

The sound was off. The building ignored the style guide. The connection was fluffy and verbose. And worst of all, location were hallucinations — the AI sometimes described Semrush features that don't exist, and successful convincing detail. 

I tried everything I could deliberation of to amended the output. Using different AI models. Tightening the prompts. Splitting drafting into smaller steps. Giving it the style guide. Giving it much past drafts arsenic examples.

None of it produced consistent, high-quality outputs. I'd get an acceptable draught once, past the adjacent tally would beryllium incorrect successful a caller way.

Eventually I stopped trying to hole the contented I was getting from n8n. The investigation half still gave america accusation for briefs the squad could constitute from, truthful we kept that moving and group the drafting aside. 

But I couldn’t extremity reasoning astir why the drafting kept failing.

It turns retired the nonaccomplishment was structural each along. n8n is awesome astatine chaining API calls — fetch this, toggle shape that, and nonstop it onward. 

Drafting an article, however, requires editorial reasoning — judgement calls astir voice, structure, and what to change. That benignant of reasoning needs to see the full article astatine once, positive reference worldly for illustration the style guideline and past examples disposable arsenic decisions get made. 

Workflow devices simply aren't built for that. 

Why I switched to Claude Code

I needed thing that could do existent editorial work, for illustration publication the original article, understand the intent down the query, and make calls astir what to alteration and what to time off alone.

I looked astatine a fewer options and kept coming backmost to Claude Code. 

Here's what made it fit:

Claude Code is an supplier that runs wrong a files connected your computer. The pipeline is that folder. The style guide, past drafts, the investigation output, and the article being updated are each files wrong it.

Claude Code sounds what it needs erstwhile it needs it, and the activity it does becomes different record the adjacent measurement tin use.

The structural quality from n8n is successful really the AI fits into the workflow. In n8n, you build the workflow successful advance, and the AI does 1 circumstantial step, for illustration penning a conception aliases summarizing data. 

In Claude Code, the AI runs the workflow itself, reference the files, deciding what to do, and penning the outputs. Combined pinch accomplishment instructions that show it what to do astatine each step, Claude Code has some the discourse drafting needs and the constraints that support it from going disconnected the rails.

That's what made the difference. 

The AI had entree to what it needed erstwhile it needed it, and a defined occupation astatine each step. The activity it produced was a record the adjacent accomplishment could prime up and a writer could unfastened later to check.

I rebuilt the full pipeline successful Claude Code, including the API calls that had been moving good successful n8n. With everything successful 1 folder, the drafting measurement could publication the investigation output, the original article, past drafts, and the style guideline whenever it needed them.

And it worked. 

The pipeline produces drafts our writers tin edit and publish, and a way of files they tin cheque erstwhile thing looks off.

Nine skills, extremity to end

The pipeline I built successful Claude Code is 9 skills, chained together by a maestro book that runs them successful order.

I springiness it the URL of the article I want to update and a target keyword, and I get backmost a draft. The draught goes done our normal editorial workflow the aforesaid arsenic immoderate different article: review, revisions, editing, and images. Our squad makes each editorial call.

Here are the 9 skills:

1. Fetch the unrecorded article
2. Research the SERP and competitors
3. Run an EDI semantic similarity cheque against our existing piece
4. Synthesize an update plan
5. Identify outdated content
6. Audit merchandise mentions
7. Draft the updates
8. Generate a side-by-side comparison of the original and the caller draft, pinch changes highlighted
9. Format the consequence for publishing

I kept it astatine 9 skills connected purpose. It was the smallest number that gave maine a chopped accomplishment for each determination the pipeline needed to make.

And 1 creation prime turned retired to beryllium really important. Every accomplishment saves its activity to a record earlier the adjacent 1 runs.

Those files are what I telephone the pipeline's artifacts. They see the research, the plan, the draft, and the side-by-side comparison. Saving each measurement arsenic a record intends immoderate azygous accomplishment tin beryllium re-run without starting over, and anyone tin unfastened the files to cheque erstwhile a draught looks off.

What changed erstwhile the Claude Code pipeline ran

Two things changed erstwhile the Claude Code pipeline started working: 

1. The hallucinations the AI still occasionally produced became easy to catch
2. The drafts started reference for illustration we wrote them

Any AI procreation measurement tin hallucinate sometimes. The pipeline is built to drawback them fast.

Dana — 1 of our contributors — was reviewing a draught and ran into plausible-looking instructions for a characteristic that doesn't exist. The benignant of correction that, successful the aged n8n version, would person either slipped done aliases costs 20 minutes of cross-checking.

She opened the side-by-side diff, looked astatine the aforesaid conception successful the original article, saw the original didn't mention the workflow, and replaced the fabrication. The full point took astir a minute.

Here’s what the diff artifact looks like:

That's what the artifacts are for. The AI is still going to make mistakes. The pipeline is built truthful a reviewer tin drawback them and cheque successful 1 infinitesimal alternatively of 20 minutes.

The bigger communicative is what happened crossed runs.

For months, I'd been trying to get the drafting measurement to nutrient thing that publication for illustration Semrush. Meaning the correct attack to voice, tone, structure, and really we picture our ain products. In n8n, I'd get a draught that possibly nailed 1 of those things and missed 3 others. And the adjacent run, I’d get a different combination. 

But successful Claude Code, 3 runs pinch mini adjustments betwixt them sewage maine there. By the third, the drafts were consistently strong.

The sound matched the existing article. The building followed our style guide. The reside was Semrush. The marque positioning was right. The AI sewage the merchandise descriptions correct. The aforesaid benignant of errors didn't support showing up successful different places.

This was the portion I hadn't expected. Months of adjustments successful n8n hadn't gotten maine here. Three runs successful Claude Code did.

Dana still caught things, but they were the smaller editorial fixes immoderate draught needs, for illustration sharpening an opening, reframing a section, aliases smoothing a clunky transition. The drafts nary longer arrived pinch the bigger problems n8n had fixed us, for illustration the incorrect voice, ignoring the style guide, aliases fabricated Semrush features.

Dana's feedback aft respective runs was that the penning was overmuch amended than what we'd produced before. And the side-by-side position was really useful.

What ended up mattering

Three things held up crossed each run.

1. Drafting needs afloat context. Treating the LLM arsenic 1 measurement successful a workflow gives you inconsistent writing. The drafting activity has to spot the article, the style guide, and the investigation astatine the aforesaid time.
2. The way of files is the system. Every accomplishment saves its activity earlier the adjacent 1 runs. That way is really our squad catches problems, and really I tin re-run immoderate azygous measurement without starting over.
3. Fewer skills, much refinement. Nine covered the work. Every clip I've been tempted to adhd a tenth skill, the correct move has been to sharpen 1 of the existing nine.

The pipeline is running, the squad is utilizing it, contributors are redeeming important time, and the feedback has been much affirmative than thing we've had pinch AI-generated content.

If you're hitting a value ceiling pinch AI content, commencement by asking wherever your AI is making its penning decisions. If they hap wrong a workflow step, that's wherever the ceiling is coming from.

Move the drafting activity location the AI tin publication your files directly. That mightiness beryllium an supplier for illustration Claude Code aliases immoderate instrumentality that gives the AI persistent entree to reference material. That's the move that collapsed done the ceiling for us.