How to build an AI content workflow from the ground up

I’ve spent months building and rebuilding an AI content pipeline in Claude Code. It now supports article updates and production for a company blog and external publications, and it usually gets pieces to about 95% of the way to publication.

That process taught me an important lesson: The hardest part isn’t getting AI to produce an article. It’s figuring out what the finished article needs to look like, then building the workflow and inputs that can reliably get you there.

If I were starting over, I’d work backward from the finished piece. I’d define what good looks like first, then determine what the system needs to take a keyword and turn it into an almost-ready-to-publish article.

Is an AI content system worth building?

For me, having this tool has allowed me to maximize my resources and create content that I otherwise couldn’t have.

However, it does come with risks. I’ve done my best to mitigate those with research, lots of human quality gates, and AI fact-checkers. But with Google aggressively noindexing noncommodity content, it may not make sense for your brand. Also, this isn’t a quick build. 

Hopefully, you already have many of the components needed to create the different agents. If not, you can build them in stages and iterate over time. Plus, you may be able to reuse some of your new agents for other workflows.

Define ‘quality’ to uncover needed inputs

A successful content pipeline generally needs to produce useful, original content in your brand voice. The pieces should be helpful to your ideal customer profile (ICP), accurately describe your business and offerings, and sound human. Ideally, they’d also have ranking and citation potential.

Once you know what you want to produce, you can create a list of everything you need to get there. Some will be constant for every run. Hard-code these into your workflow because they represent shared context required across all runs. Others, like the topic, angle, and keyword, will be input with every run.

Constants include:

• Brand explainer that states who you are, along with information about your ICP. For B2B, include industry, seniority, and pain points. The latter can be derived from client call transcripts or sales call transcripts. For B2C, include age, sex, profession, pain points, etc.
• Brand voice guidelines detailing how you communicate. This would include examples and not just adjectives. If your current brand voice guidelines are only a list of adjectives, you’ll want to expand them before using them for a content pipeline. You don’t want to say “friendly, but formal.” Instead, provide examples of what this looks like in practice, as well as examples of what to avoid. Don’t have a voice guide? Work with your LLM to create a doc based on your best content.
• Example content briefs, outlines, and articles that follow your best practices.
• Product, service, and/or methodology descriptions to help the LLM describe what you do accurately. Sales collateral could be helpful here.
• Information about your existing content. If you want internal link suggestions or to make sure that the pieces are distinct from the things that are already live on your site, you’ll want to share a Screaming Frog export with this information or your sitemap.
• Links to or copies of internal research and case studies. This will help you create noncommodity content.
• If you want your piece to be cited or if you want it to rank well, you also need to include information about best practices. Your organization may already have these types of writing or publication guidelines. If not, you can use Claude to help you research and generate this document. At its most basic, you want it to create a meta description, suggest a URL slug, and integrate any keywords or concepts that you give it in a natural way. You can make this more robust by requiring SERP research to analyze top-ranking or top-cited content for particular terms, requiring answer-forward passages, keeping paragraphs a particular length, etc.

You may have more criteria for a finished piece. If so, list them out.

Dig deeper: How to train Claude to sound like your brand

Determine the order of operations

Now that we know what quality looks like and what inputs it takes to get there, we need to determine the order of operations.

To some extent, this will mirror a regular content process. After all, that’s what we’re replicating. So while you can think of this build as onboarding a new writer and editor, you also need to consider the complications that using AI to create content introduces. The two main issues are robotic-sounding language and hallucinated “facts.”

Generally, with content creation, you research a topic, create the outline, write the piece, and edit it. Sometimes, different people handle these steps depending on the size of the team. 

The agents in your workflow will handle those same tasks. However, you’ll also want to build in additional human review gates that you might not need with a human-run content workflow.

To keep everything flowing appropriately and make sure no steps are skipped, I recommend an orchestrator agent that describes the workflow from start to finish and what each agent is responsible for. This document will need to be updated if you change your workflow or agent responsibilities. The orchestrator agent can be created at the end or simultaneously.

For each of the following steps, I’ve shared suggested documentation and information about how to build it. Work with your favorite LLM to build the actual agent docs, but be sure to review them so that you understand what they do and why. This will make troubleshooting easier.

Step 1: Workflow kickoff

Consider how you want to kick off your workflow. I have a separate locally hosted dashboard that lets me input a keyword and an angle. Once I click submit, Claude starts researching the topic.

If you have multiple ICPs or a specific product you want to mention per article, specify those as initial inputs before your workflow kicks off.

To start, I recommend focusing on one type of content, such as blog posts or LinkedIn posts. Later, you can add if/then functionality to create multiple content types using the workflow. 

It’s best, though, to get a working process for one content type and then add on, rather than trying to do too much at once and ending up with a workflow that doesn’t work well for anything.

Step 2: Research

Once a topic is submitted, it should start an agent that researches the topic, what your brand’s already written about it, and current SERPs to identify gaps your new piece could fill. I have mine output a dossier that can be handed off to subsequent agents.

• Agent: Researcher
• Suggested documentation:
  • List of reputable, citable sources in your industry where the agent should look first.
  • List of sources to avoid.
  • Research requirements, such as timeliness or size thresholds.
  • Your sitemap or a list of your content.
  • Links to or documentation for your original research and data.

• If you want your agent to look at SERPs, note what types of results it should analyze. For example:
  • Do you want it to look at AI Overviews to see who’s being cited for a particular keyword? 
  • Do you want it to look at top-ranking sites except for list sites? 
• How to build:
  • Share with your LLM precisely what you want the research to achieve. 
  • Be as specific as possible, especially when it comes to what qualifies as a good source and your own content.

Dig deeper: How to build a Claude Code-powered second brain for agency work

Step 3: Outline

Once the research is complete, I recommend having Claude generate an outline and adding a human review gate. 

That way, you can get a sense of what the finished piece will look like without having expended as many tokens. At this point, you’ll need to decide whether to scrap the piece, revise the outline, or continue.

• Agent: Outliner
• Suggested inputs and documentation: Example outline, voice guide, research dossier from your research agent
• How to build:
  • Provide the LLM with an example outline and detail all the components that it must have. For example, if you like hierarchical outlines, state that clearly and describe the must-have sections.

Step 4: Write

The LLM can now create the content based on the outline and research.

• Agent: Writer
• Suggested inputs and documentation: Research dossier, outline, brand voice guide, ICP information, case studies, first-party research
• How to build:
  • Provide the LLM with an example of high-quality content written from an outline. 
  • Have it analyze the narrative flow and logical progression of the piece, as well as the words used and how they’re used. If you follow any rules, like bottom line up front (BLUF) or mutually exclusive, collectively exhaustive (MECE), be explicit about when to apply them.

Dig deeper: 6 content audit workflows to build in Claude

Step 5: Edit, fact-check, and iterate

Rather than placing a human review gate after content is written, I like to run at least one editor pass.

For a robust system, I use a regular editor who checks for conformity with brand standards and assesses the draft against the stated angle and outline. I also run a fact-checker that verifies accuracy by double-checking stats and an AI editor to help remove AI tells.

These editors don’t just make copyedits. They also assess whether the piece needs additional revisions and suggest them. For better results, each of these should be run in a new context window.

I originally had one editor handling structure, coverage, and style together. Splitting it into two — one for structure and coverage and another for phrasing and AI tells — produced better output than asking one context to fix everything at once.

Similarly, I found that asking an editor to fact-check resulted in two poorly done jobs. It’s also a great place to put in a feedback loop.

• Agents: Editor, fact checker, AI editor.
• Suggested inputs and documentation: Editorial guidelines, voice guide, research dossier, writer’s draft.
• How to build: Each of these three agents needs distinct instructions.
  • Editor: The initial editor should be responsible for verifying style guide compliance. If you have specific words you don’t want used, it should check for those and remove or replace them. It can also check for narrative structure, logical order, vague phrasing, and any missing bridges between sections. Be clear on what its job is. Provide examples as necessary.
  • Fact checker: The fact-checker exists to help remove hallucinated information. It should have an adversarial bent. It should assume that everything stated in the article is wrong and attempt to disprove it.
  • AI editor: No doubt, there are certain AI tells that raise your hackles. Here’s the chance to build an agent to strike those from your prose. In addition to your own personal pet peeves, consider adding some of these signs of AI writing to your agent doc. This editor should not focus on style, but strictly just on the AI tells.

This is the final human review gate, unless you want to publish to a CMS via API. I run two passes with my editors before a piece gets handed to a human. This reduces the amount of editing humans need to do, but it won’t get the piece to 100%. Nor should you ever want to publish something that hasn’t been touched by a human.

Give your AI content system the right context

Opus and Fable can help you create these agent docs quickly. However, you still need to provide the context that will make the content worth reading.

If you don’t have the appropriate inputs, like information about your ICP, examples of quality content, or links to first-party research, no content workflow is going to help you create noncommodity content that resonates with your audience.

Give your favorite LLM this article and have it work with you through the process so you have all the information you need as you build.

Dig deeper: How to scale SEO content updates with Claude Code

Where to take your content workflow next

If you want to take this further, you can:

• Add keyword and entity research to support the research step.
• Include additional content types, either as standalone pieces or as add-ons for the pieces you’re creating, like a LinkedIn newsletter based on your blog post.
• Integrate a blog content refresh process. 

The content refresh process has been an important part of my workflows, allowing me to update content with new stats and information. If the entire workflow feels like too much, remember that you can start with one agent.