Despite what I think of myself, I am but a mere mortal. One person, with a lot of thoughts, a lot of ideas, and not a lot of capacity.

And because I’m a masochist, I built, and run my Siren, which is a ludicrously powerful incentive program plugin, but honestly I look at it as incentive payout infrastructure. The marketing challenge is that it’s very hard to explain what it actually does, because the real value proposition doesn’t apply with something like Siren until it’s put up against a problem that it is capable of solving.

There are hundreds of conversations to be had about Siren, because there are dozens of use cases under it. Affiliate programs. Employee bonus programs and SPIFFs. Refer-a-friend programs. Speedy-rewards-style point programs. Hell, a car dealership could plausibly run its entire commission structure on it. Each is a different audience, and each one could easily be months worth of content writing.

I guess my point is, I’ve determined that the only way to solve my communication problem with Siren is with a ton of content. I can’t have every conversation personally. But I can multiply myself with AI, if I do it without shipping slop.

This workflow is for content that has to land. Pillar posts. Sales pages. Blog posts you want a real person to read and remember. Content that you share and actually think “this is really going to help someone”.

I’m no stranger to writing blog content consistently. I’ve been talking about it since 2016 and have been writing since 2009. But with the advent of AI, and the truly grotesque amount of content that Siren requires, I’ve really been shaking things up, and while it’s still a lot of work, I’m consistently publishing several blog posts a week right now.

And no, it’s not generic AI slop. It’s real content, with a real perspective, and something that’s actually worth saying.

Wanna see how I do it? Thought you might. Let’s get into it.

My content creation workflow

The content creation workflow I use consists of these ingredients, in order:

1. A voice profile built once and reused everywhere
2. An interview with me, to extract my opinion, perspective, and what I have to say
3. A multi-pass content pipeline with specialized agents that don’t share history
4. A customer avatar review for cornerstone pieces
5. The human pass, where I finally read what was written, and (gasp) actually write some of the content

The voice profile

The voice profile is basically a document that describes your writing style (or styles, if you have different personas based on where you’re at). I built mine by analyzing my own published essays. Not what I thought I sounded like. What I actually wrote, across years of posts on this site. I used AI to help identify patterns and things.

The result is a roughly 1,500-word document covering the patterns. How I open and close. How I use italics (constantly, on the word that carries the argument). The words I’d never use. Signature phrases that become tells the moment they get stacked. My proclivity for self-deprecating humor, the tasteful use of cussing. Sentence rhythm rules. Even the way I drop in parenthetical asides, and my habit of droppin’ the “ing” on words to better match how I speak in real life.

It goes into every prompt I run. Without it, the AI averages me toward “generic blog voice.” With it, the AI has a target to aim at.

A voice profile by itself isn’t enough, though. Honestly, it doesn’t even really make the content good, and the AI will still fail to really nail your voice, but it still gets it a lot closer for the editing phase later.

Without the next four steps, however, the voice profile is just a costume the AI puts on while it still says nothing of substance.

The Interview

I work conversationally. This is my natural state. I want to talk. I’d rather dictate a five-minute discursive answer than outline a piece in bullet points before I’ve even thought about it. When I’m talking, I walk. When I walk, my blood gets flowin’, which helps me think better, while also getting my ass away from my desk.

So my workflow accommodates that. After the voice profile is loaded, before any drafting starts, I run an interview. I instruct the AI to be a reporter, asking me questions about the topic we’re going to write about, one at a time. I answer out loud. Sometimes the answer is thirty seconds. Sometimes it’s a five-minute stream of consciousness that wanders into a tangent and comes back. Then it asks the next question based on what I said.

This continues for about 20 minutes or so, usually, and the AI plays back what it heard and I correct what didn’t land. By the end, the AI has my actual thinking. Not a guess. The thinking, in my own words, with the parts I emphasized and the parts I shrugged at.

This matters because all the model does is predict. Without your thinking loaded in, it picks the most-likely next word, then the most-likely sentence after that, and so on down to the most-likely paragraph. That output is what anybody with a similar prompt would have gotten. It’s the slop default. This is what I keep coming back to in AI doesn’t have your perspective. The interview is what gets your thinking onto the page before any sentence of the draft exists.

Not only does the AI have this context, but I actually got the context out of my head, and sometimes that context isn’t obvious when I get started. This post, for example, started as a social post I shared. I knew I had opinions, and thought there may be an opportunity to write something about it, so I shared the post with my AI agent, and by the end of the interview I had a whole dang cluster of content to publish. In other words, I literally didn’t even know what we were writing when we started, but the content presented itself by the end of the interview.

The multi-pass content pipeline

Once we know what we’re going to write, I run the agent through a content pipeline. This is a multi-phase process, where the agent writes, and cleans up the content.

1. The first pass is brief assembly, where a research agent gathers sources, locates the cross-links, and structures the argument before any prose gets written.
2. The second pass is section-by-section drafting against that brief, with the voice profile and the interview transcript both loaded.
3. The third pass is a voice and quality refinement that scans for em-dashes, audits signature-phrase usage, and verifies italics land on argument-carrying words.
4. The fourth pass is the structural audit, which catches the deeper tells, the bold-label paragraphs, the enumeration disguised as prose, the identical section templates that signal “AI generated” even when no individual word does.

Each pass is a separate agent with its own job, and none of them know what the others did, which I think is the most important detail here. Because the agents work in isolation, the editors work much more reliably because their context isn’t polluted with previous information that they have to filter out.

Customer avatar review

This step involves a set of simulations where I run a collection of different avatars to review the content themselves, and synthesize their thoughts on the content, and what they did from there (click through, bounce, etc).

This is hardly perfect. After all, AI is still essentially pretending to be that role, but those avatars read the content with different biases, and they often catch things that neither me nor the agent could catch without being that avatar.

Picture an affiliate manager at a mid-size SaaS company, three years into the job, tired of click-only attribution arguments with finance, scrolling a blog post on her phone between Slack messages. That’s a customer avatar. I have a handful of them, and for high-stakes content I have AI agents that act as those personas, read the post, and tell me what their experience was.

Others in the rotation include a WordPress agency owner thinking about adding affiliate revenue, and a solo course creator who’s never run a partner program but wants to. Each reads with their own bias. Each tells me something different about what landed.

I run this set of avatar reviews, and then re-run the content pipeline to patch up the observations and feedback we received, and do so until I’m satisfied with the avatar’s opinion of the content.

The human pass

This is the first step in the process where I even look at the content. By the time I’m reading the draft, the AI has been through four or five passes, and the content should match my voice, answer and solve the questions that the various avatars had.

What’s funny is because I’ve spent that time being interviewed, I have a good sense of exactly what the wrong-feeling sentences are. They practically jump off the screen, because I already decided what the post should say back during the interview. I know the shape of the argument. So when a paragraph wanders into a take I don’t hold, or softens a position I felt strongly about, the wrongness is loud. The edit pass is relatively fast, and surgical.

This is the moment that makes the difference between AI-assisted content and AI slop. Skipping this pass produces slop. Including it makes the post mine. Even with all of the steps I listed above, the difference between what the workflow produces and what I end up with after spending 20 minutes going through and editing the content is genuinely night and day.

Conclusion

To multiply yourself with AI without ending up in the AI-blind pile, build a voice profile, interview yourself before you draft, run multiple passes with separate agents, and read the draft last with the standard you set during the interview still loud in your head. The order is the moat. The system is what lets a mere mortal keep up with a product that needs hundreds of conversations.