How to Use Claude Code + AgentKit for SEO Content (2026)

by Jasmine, a developer who writes SEO content with Claude Code most working days.

Most "AI for SEO" guides show you how to analyze data or optimize a page that already exists. Almost none show the part I actually spend my time on: turning a keyword into a finished, publishable article. That is the whole point of this guide - the complete writing loop, with the real prompts I use, honest notes on what breaks, and where a ready-made skill saves you an afternoon of setup.

Which AgentKit? (30-second disambiguation)

Before we go further, one thing trips up every search for this topic. "AgentKit" now means two unrelated products:

This guide uses AgentKit for Claude Code - the kit at agentkit.best, installed via the ak CLI, which ships an seo skill and an ak-copywriting skill. It is not OpenAI's AgentKit (Agent Builder plus ChatKit, launched 2025-10-06), which is a different platform for building agents. There is also a free name-twin skill on GitHub (AgriciDaniel/claude-seo). That one is a single community script; the AgentKit skills I use are a maintained pair with updates. I mention this so you don't install the wrong thing and wonder why the commands differ.

Why use Claude Code for SEO content (vs a chat window)?

You can absolutely write SEO drafts in a chat window. I did it for a year. The problem is that a chat has no memory of your files and no hands. Every article is a fresh copy-paste ritual: paste the brand voice, paste the competitor notes, paste the outline, copy the draft back out, reformat it, and do it all again next week.

Claude Code is a terminal agent from Anthropic that runs inside your project folder. That changes the shape of the work in three concrete ways:

• It reads and writes real files. Your brand voice lives in a CLAUDE.md file, your briefs live in /briefs, your drafts land as .md or .html. The agent opens them, edits in place, and saves. No copy-paste round trip.
• It runs commands. It can grep your existing posts for internal-link targets, count characters in a meta description, or check that no two articles target the same keyword.
• It is repeatable. Because the instructions live in files and skills, the same pipeline runs identically for article #1 and article #40. A chat prompt you retype is not a workflow; a skill you invoke is.

If you are brand new to the tool, I wrote a plain-English primer on what Claude Code is and how it differs from an editor plugin. The short version: it is closer to a junior teammate you brief with files than to an autocomplete box. That framing is what makes the SEO pipeline below possible.

The honest trade-off: a chat window is faster for a one-off paragraph. Claude Code wins the moment you write content regularly and want the same quality bar every time.

What you need before you start

Five things. You can start with the first three and add the rest later.

• Claude Code installed and signed in. A paid Claude plan (Pro is $20/month; Max tiers are $100 and $200/month) makes long article sessions comfortable; there is a limited free tier to test the flow.
• An SEO + copywriting skill. Either roll your own (a folder of instructions the agent loads) or use a maintained one. If you are new to the concept, here is how Claude Code skills work - they are reusable instruction packs the agent activates on demand. AgentKit ships an seo skill and ak-copywriting skill pre-wired for this exact pipeline; more on that later.
• A CLAUDE.md with your brand voice and rules. This is the single highest-leverage file. Put your tone, banned phrases, target reading level, meta-title and meta-description length limits, and internal-link conventions here. Every session reads it automatically.
• Keyword data. Claude Code cannot browse live search results, so you feed it the data: target keyword, a few competitor URLs, and any metrics you have from Google Search Console, Semrush, Ahrefs, or a manual SERP look.
• A content folder structure. Something as simple as /briefs, /drafts, and /published is enough to keep the agent oriented.

The full workflow - from keyword to published article

This is the backbone: six steps, each with a real prompt or command, what the agent gives back, and where a screenshot belongs. I run these as a sequence, saving each output to a file so the next step can read it.

Step 1 - Research the keyword & search intent

Claude Code has no live SERP access, so I hand it the raw material. I paste the target keyword and the top competitor URLs (I copy the visible page text into a file, or use a fetch/scrape step if I have one wired up), then ask it to do the analysis I would do by hand.

Target keyword: "claude code for seo content"
I've saved the top 5 competitor pages to /research/competitors.md.

Read that file and give me:
1. The dominant search intent (informational / commercial / transactional)
2. Every H2 each competitor uses, grouped by theme
3. Topics ALL of them cover (table stakes)
4. Topics NONE of them cover well (my opening)
5. Rough word count of the top 3

What comes back is a structured intent classification plus a gap list. The gap list is the gold - it tells me where I can add information the ranking pages are missing, which is exactly how you earn a spot on a mature SERP instead of writing the ninth identical guide.

Step 2 - Generate the content brief

Next I turn that research into a brief the drafting step can follow blindly. I ask for per-section word counts, keyword placement, the competitor gaps mapped to specific H2s, and a schema plan.

Using /research/intent.md, write a content brief to /briefs/brief.md with:
- H1 + meta title (≤60 chars) + meta description (130-150 chars)
- A full H2/H3 outline with a word-count target per section
- Where the primary + 2 secondary keywords go
- Which competitor gap each section fills
- The schema to add (FAQPage if there's an FAQ)
Keep every claim checkable - mark anything you're unsure of as.

A small meta note that builds real credibility: the article you are reading was written from a brief structured exactly like this - per-section word counts, competitor gaps, keyword placement, a schema plan. The pipeline is not theoretical; it is producing this page.

Step 3 - Draft the article (anti-slop)

This is the step everyone gets wrong. Point an AI at a brief and say "write it" and you get slop: hollow intros, "in today's fast-paced world," repeated ideas, confident fake statistics. The fix is a hard, specific prompt plus the brand voice from your CLAUDE.md.

Here is the anti-slop prompt I actually use:

Draft /drafts/article.md from /briefs/brief.md. Follow CLAUDE.md for voice.
Rules:
- No filler openings, no "in today's world", no throat-clearing. First
 sentence answers the question.
- Every claim is either checkable or marked. Invent NO statistics,
 prices, or study citations.
- Include at least one concrete, real example per major section.
- State limitations and trade-offs honestly - don't oversell.
- Vary sentence length. No paragraph over 4 sentences.
- Where I should add first-hand experience or a screenshot, leave a marker instead of faking it.

The two lines that matter most are "invent no statistics" and "leave a marker instead of faking it." Together they turn the model from a confident fabricator into a careful drafter that flags its own gaps. You will still edit heavily - but you will be editing a real skeleton, not deleting slop.

Step 4 - On-page optimization

Because the draft is a file, I can optimize on-page elements in the same session instead of pasting into a separate tool. I ask for everything in one pass and - critically - I make the agent count characters, because models are notoriously bad at length limits unless told to verify.

Optimize /drafts/article.md for on-page SEO:
- Meta title ≤60 chars - print the character count next to it
- Meta description 130-150 chars - print the count
- Check heading hierarchy (one H1, logical H2/H3)
- Suggest 3-5 internal links from /published (grep for related topics)
- Write descriptive alt text for each image placeholder
- Draft an FAQ (5-6 Q&A) and the matching FAQPage JSON-LD

The "print the character count" instruction is the difference between a 68-character title that gets truncated in results and a clean 54-character one. Trust the count, then eyeball it yourself.

Doing on-page in the same session has a second benefit: the internal-link suggestions are grounded in files that actually exist. When I ask the agent to grep /published for related topics, it proposes links to real, live URLs on my site instead of inventing plausible-looking slugs - a mistake that is easy to miss when you optimize in a separate tool and paste links back by hand.

Step 5 - Human edit, fact-check & E-E-A-T

Non-negotiable. This is where the article stops being AI output and becomes yours. I do four things:

• Resolve every marker. Open the source, confirm the number or date, replace the marker. If I cannot verify it, it comes out.
• Fill every marker with a real screenshot or a genuine first-hand note - not a plausible-sounding invention.
• Add experience the model cannot have: what actually broke, what surprised me, the trade-off I only know from doing it.
• Cite and date. For any capability or launch claim I link the official source with the date - Anthropic's docs for Claude Code features, or the vendor page for a product fact.

Google's own guidance is that content should demonstrate experience, expertise, authoritativeness, and trust; see Google Search Essentials on helpful, people-first content (accessed 08/2026). An AI first draft has none of that by default. You inject it here, or the article stays generic.

Step 6 - Publish & track

The last step depends on your stack. I write to Markdown or HTML, so publishing is a push to the CMS - Strapi in my case, but the same file-based flow suits WordPress (via an importer) or a static site. Then:

• Submit or let the sitemap ping the search engines.
• Log the target keyword and publish date so you can measure the article later.
• Check back in 2-4 weeks for rankings and any AI Overview appearance, and feed what you learn back into the next brief.

If you want to run this whole loop as a repeatable, agent-driven process - briefs, drafts, and review as defined steps rather than ad-hoc prompts - I break that pattern down in my guide to the AI dev workflow with AgentKit. The same brainstorm-plan-build-ship rhythm that works for code works for content.

Do it faster - AgentKit's ready-made SEO + copywriting skills

Everything above works with skills you build yourself. The catch is that building them well - the brief format, the anti-slop rules, the on-page checks, the schema plan - takes real time and iteration. If you would rather skip the assembly, this is where a maintained kit earns its keep.

AgentKit ships an seo skill and an ak-copywriting skill (part of its Marketing kit) that come pre-wired with a pipeline like the one in this article - content briefs with per-section word counts, competitor scoring, keyword-placement guidance, and conversion-copy formulas. Setup runs through the ak CLI in under five minutes, and the skills get lifetime updates so the workflow keeps pace as the tools change. It is the difference between a single free github script and a maintained pair that stays current.

Want the pipeline without the setup? The Marketing kit (which includes the seo + ak-copywriting skills) is $99, and the site does not list a recurring fee for kits; the Engineer + Marketing bundle is $149. There is a money-back guarantee (the site does not state a fixed duration). For the full picture, read my complete AgentKit review or see AgentKit pricing (20% off via link) and re-check the live numbers before you buy.

To be clear about what you are paying for: the skills do not remove Step 5. You still fact-check and add experience. They remove Steps 1-4's setup, which is the tedious part.

Real example - one article, end to end

Let me show the pipeline on a real piece rather than describe it in the abstract. I wrote a guide on internal linking using exactly the six steps above.

Step 1-2 (research + brief): I fed in five competitor URLs. The gap analysis flagged that every ranking page explained what internal links are but none showed a repeatable audit of an existing site. That became my angle, and the brief built the outline around it.

Step 3 (draft): With the anti-slop prompt, the first draft came back with three markers on statistics it refused to invent and two markers where a screenshot belonged. That refusal is the behavior you want - it is honest about its blind spots.

Step 4 (on-page): The initial meta title came back at 63 characters. Because I had asked for a printed count, the overrun was obvious and I trimmed it to 55 in one edit instead of finding out weeks later from a truncated snippet.

Step 5 (the honest part): This is where I spent the most time. The draft's "why internal links matter" section was technically correct but generic - pure textbook. I rewrote it with a specific mistake I had made (orphaned pages after a site migration) that no model could have known. I also cut an entire subsection the model padded in that added nothing. Roughly 40% of the finished article was human editing. That is normal, and anyone who tells you AI content ships at 95% is selling something.

Result: a publishable article in roughly a third of the time the fully manual version would have taken, with the quality gate intact because a human held the last step. If you want to run this same loop with the skills already assembled, you can try the ready-made workflow and adapt the briefs to your own voice.

Best practices - automate vs keep human

The single most useful mental model here is a clear line between what the agent does well and what must stay with you. Cross that line and quality drops fast.

Rule of thumb: automate the mechanical and repetitive; keep the judgment and the experience. The value you add as a writer is precisely the column on the right - so protect it.

Honest limitations

If a guide only sells you on the upside, distrust it. Here is what genuinely goes wrong:

• No live SERP access. Claude Code cannot see current search results. Everything it "knows" about competitors is what you feed it. Stale or thin input equals a weak brief.
• Hallucinated stats and citations. Without a strict prompt, it will confidently invent a statistic or a study. The anti-slop rules reduce this; they do not eliminate it. You must verify every number.
• Slop by default. The baseline draft trends generic. Good output is a function of good prompting plus a strong CLAUDE.md, not the model alone.
• Cost and limits. Long article sessions consume plan usage or API tokens. Heavy content teams should budget for a Max plan or watch API spend (Opus and Sonnet are billed per million tokens).
• Verification time is real. The pipeline speeds up drafting, not fact-checking. Step 5 does not compress. If anything, doing it properly is the honest tax on using AI at all.

None of this makes the tool a bad idea. It makes it a tool - one that rewards a skilled operator and punishes a lazy one.

Frequently asked questions (FAQ)

Can Claude Code actually write SEO content?

Yes, when you drive it as a file-based pipeline: research the keyword, generate a brief, draft with anti-slop rules, run on-page optimization, then human-edit. It handles research synthesis, drafting, and on-page mechanics well. It cannot replace fact-checking or first-hand experience, so a human still owns the final step.

Is AI-written content against Google's guidelines?

No. Google's guidance targets low-quality, unhelpful content regardless of how it was produced - not AI use itself. Content that demonstrates real experience, expertise, and accuracy can rank whether AI helped write it or not. The human editing and fact-checking steps are what keep it on the right side of that line.

Do I need a paid plan or a paid kit?

You need a Claude plan (a limited free tier exists; Pro at $20/month is comfortable for regular writing). A paid kit is optional - you can build your own SEO and copywriting skills for free. A kit like AgentKit just saves the setup time by shipping those skills pre-wired.

Claude Code vs ChatGPT for SEO content - which is better?

ChatGPT (a chat window) is faster for a one-off paragraph. Claude Code wins for repeatable, file-based production because it reads and writes your actual content files, runs commands, and can be driven by a reusable skill - so the same quality bar applies to every article, not just the one in front of you.

Which AgentKit does this guide use?

AgentKit for Claude Code (agentkit.best, CLI ak), which ships the seo and ak-copywriting skills. This is not OpenAI's AgentKit (Agent Builder and ChatKit), which is an unrelated agent-building platform launched in October 2025.

Can it publish directly to WordPress or Strapi?

The workflow produces Markdown or HTML files, which slot straight into a static site or Strapi, and into WordPress through an importer or the REST API. Claude Code does not push to your CMS automatically by default, but because the output is a plain file, the publish step is a simple push or import.

Conclusion & next steps

The whole method fits in one loop: research the keyword, generate a brief, draft with anti-slop rules, optimize on-page, human-edit and fact-check, then publish and track. The magic is not the model - it is treating content as files and keeping the last step human. Build the skills yourself, or skip the setup with a maintained kit and spend your time where it counts.

New here? Start with what Claude Code is, then read the full AgentKit review and the AI dev workflow guide to run this as a repeatable system.