AI & Development
The exact workflow we use at Stika Studio to go from a content idea to a published, SEO-optimized blog post using Claude agents
Building an AI content workflow that works in production requires more than choosing the right AI model. It requires defining exactly which steps are human and which are AI, what format the output of each step takes, and where quality gates sit to catch errors before they propagate downstream. Here is the exact workflow we run at Stika Studio - every step documented, with the reasoning behind each decision.
Every article starts with a human brief. The brief specifies: the target keyword (what someone would search to find this article), the search intent (are they trying to learn something, compare options, or take action), the unique angle (what this article has that no existing article on this topic has), and any specific requirements (word count targets, internal links to include, external sources to reference).
The brief is the part that cannot be automated. Deciding which topics to write about - based on keyword research, audience needs, competitive gaps, and strategic priorities - requires judgment that depends on business context the AI cannot access. The brief is where that judgment is captured and handed to the pipeline.
A good brief is a half-page document. A bad brief is a sentence. The difference in output quality between "write about AI agents" and a structured brief specifying the target keyword, search intent, angle, and format is the difference between an article that ranks and one that does not.
Before writing, we check the current SERP for the target keyword. The blog-writing agent uses the WebSearch tool to retrieve the top 10 results and notes: the dominant content format (guides, listicles, comparisons, product pages), the average article length, which questions appear in "People Also Ask", and whether a Google AI Overview appears for the query.
This step informs the content format decision. If the top 10 results are all step-by-step guides, the article should be a step-by-step guide. If they are all comparison tables, the article should lead with a comparison. Format matching is mandatory for ranking; format mismatches produce articles that compete with the wrong content type and do not rank regardless of quality.
The blog-writing agent takes the brief and SERP research and produces a full draft in the correct technical format. For our React-based blog, that means a JavaScript file with the correct data structure: slug, SEO fields (title, description, ogImage), tags, category, reading time, body HTML, and CTA block.
The agent's system prompt encodes all format requirements explicitly. The output is not a Word document or Markdown file that then needs reformatting - it is the exact file the system needs, already in the src/data/posts/[slug].js format. This eliminates the reformatting step that consumed significant time in our earlier, less structured approach.
Draft generation takes 5 to 10 minutes for a 1,200-word post. The draft covers all planned sections, includes internal links where specified in the brief, and meets the minimum quality gates encoded in the agent's system prompt (word count, keyword placement, no AI filler phrases).
The draft is reviewed by a human before any SEO processing or technical implementation. The review is focused on three things: factual accuracy (are all specific claims correct?), brand voice alignment (does this sound like us?), and angle execution (does the article deliver on the brief's unique angle?).
Most AI-generated drafts in our pipeline pass this review with minor edits - word choices, one or two factual corrections, the occasional structural adjustment. When a draft requires significant rework, it is almost always because the brief was vague or because the unique angle was not specified clearly enough. Bad output from a well-specified agent traces back to a bad brief more often than to a failure in the agent's execution.
After human review and any edits, the SEO agent validates the article against a defined checklist:
echo -n "title" | wc -c)The SEO agent produces a structured handoff block for the technical implementation step, listing what schema types are needed and what the sitemap.xml update should contain.
The technical SEO agent executes the handoff: adding the post to src/data/posts/index.js, updating public/sitemap.xml with the new URL, and verifying that the post renders correctly. This step is fully mechanical - there are no judgment calls, only defined specifications to implement.
The deployment step runs npm run build and firebase deploy --only hosting. This step is fully automated, triggered by confirming the technical implementation is complete. The build process validates that all imports are correct and the JavaScript compiles without errors - catching any syntax errors introduced in earlier steps.
Brief: 10 minutes (human). SERP research: 5 minutes (AI). Draft: 10 minutes (AI). Human review: 15 minutes (human). SEO validation: 3 minutes (AI). Technical implementation: 5 minutes (AI). Deploy: 3 minutes (automated).
Total: approximately 51 minutes per article, of which 25 minutes is human time. The throughput for a single developer working with this pipeline is 4 to 6 articles per day - roughly 10× the throughput of a manually-written workflow without AI assistance.
The workflow does not automate topic selection (which topics to write about), the unique differentiating insight (the element only a human expert could provide), or factual accuracy verification (checking that specific claims are correct). These three elements remain human because they require judgment, genuine expertise, and access to information that is not in the AI's training data or retrievable from the web.