Writing as a Perfectionist
I have high quality standards. Probably high enough that it qualifies as perfectionism. And definitely high enough to keep me from shipping anywhere near as much "good enough" content as I could.
I realized years ago that trying to edit while drafting was a sure-fire way to get yourself blocked. So I very intentionally split my writing into smaller roles: outline, research, draft, edit, publish. Very distinct roles with very distinct requirements.
The trouble was, I just moved the obstacle instead of removing it.
Rather than just spinning in circles writing and rewriting the same few sentences, I now wrote a bunch of drafts and then spun on transitions and editing the chunks together -- and never felt like it was done. I had more words, but just as few finished works.
Somewhen along the way I realized I am much more pressure prompted than early start. So I used deadlines to get things out the door. They were not to the level of quality I actually wanted, but they were "done."
The trouble with using deadlines, besides not feeling like it reflected my best work, was the physical toll of running on cortisol and spite. I fully acknowledge I'm a steam engine and need some amount of pressure to get moving and keep moving. I have just gotten better at finding the right types of deadlines to work towards and keeping an eye on the pressure gauges to stay in the safe zone.
And I say "finding" deadlines because artificial deadlines never seemed to work. My brain says "nope, that's just meant to get me to take action before I'm ready" so I rebel! (An awful lot like when people try to use artificial scarcity or manufactured pressure to sell products or services.)
Getting to Quality Content
I've been using a variation on an "interview" approach in Claude (and Gemini and ...) for quite a while. I built on the RTFM pattern I picked up early on and refined the "ask me until" format to have the LLM guide the conversation. It let me brain dump and chase tangents and add any content that felt like it might be relevant -- in ways that we don't feel great about doing to a human listener. I didn't feel like I was making someone work to follow what I was saying as I figured out what all I had to say.
I found a sustainable cadence to use the LLM as a thought partner to draw out my ideas and words rather than pulling the slot machine handle and trying to use whatever mixed up stuff the AI spat out. We've all seen the typical AI patterns in content. There are lists online like "7 signs your content was written by AI" that include hints on how to "humanize" it. We've all seen more emdashes than I use naturally -- plus the sycophantic phrases and ideas that are statistically likely, which just makes the whole piece a race to the middle for mediocrity.
I also found from early on that if I let the AI run without being in the loop, it picked things that didn't suit my preferences and needs. We had a multi-step "proto-agent" years before we saw the sorts of agents people talk about today. I refused to keep using it. Every time I did, it made choices I disliked, then made follow-up decisions that compounded the problems, and by the time it reached the end I wanted nothing to do with that "solution." It was a fun toy idea back when chat interfaces were novel, but I couldn't trust it to generate anything useful.
When I built my AI writing software, I fought back against a lot of that. Human influence at each step. Co-creating outlines. Having the AI expand to a sentence outline for the human to edit and influence before the AI drafted the text. Multiple points in the process to steer the direction and give the AI better info to guide it. And I still didn't use that -- better quality, but too many details for most folks to manage with very little context until the end. It just isn't a great way for people to actually write and capture their thoughts and feelings and worldviews.
Instead, I have had hundreds of these sorts of "ask me until" conversations that may span days or weeks as I go back and pick up where we left off. And while I could extract the decisions and reasoning that was unearthed with a single prompt, I still ended up with too many great conversations that were just languishing in the chat platform and never making it into action.
From Chat to File to Published
When I heard about Hermes Agent, I realized it might bridge that gap. Hermes Agent is a way to connect the Large Language Models (LLMs) behind the chat interfaces with your own data and files. They saw what made OpenClaw so appealing, like being able to interact from your existing apps like Slack, Discord, or Telegram, but they went beyond that to make it self-improving.
Since it is running in a real computer environment you control, Hermes Agent can push things from a conversation straight to a file on that computer. And that computer can be a sandboxed account, container, or VPS, not your main system. That was enough to push me past the hesitations I had around OpenClaw and Claude CoWork having full access to my computer.
As I started capturing ideas, I decided to mock up a site around these files. Might as well go straight to the end goal and get it ready for using in a site. Somehow my brain made it as far as a local dev version of the site and stopped.
Then I was working on a page when the family had to go somewhere. We got loaded into the car and since my wife prefers to drive, I pulled out my phone and kept working on the site design. A few minutes later I realized I couldn't see whether it was doing what I wanted without being at my computer or publishing it to the web somewhere. But I hadn't set that up, so I figured I was out of luck.
But here's where the whole "agent running on my computer but accessible via Slack" thing really came into play. I knew I had another project with all the right settings that my Hermes Agent could access and use as an example, so I tried it out. Next thing I know, it built the site, pushed it to my Cloudflare account, and I have it open in the browser on my phone.
That's the sort of "oh, wow" that got me thinking differently about using the current batch of AI tools.
The Content Interview Process
So I circled back and started to turn the process from all those Claude chats into something that I can trigger in my agent. And since I know myself and I can't just work on one site at a time, I made it a general skill.
I specified the starting info: I know it will need to know which project and the topic for the "content interview," so I have it ask for them if I don't specify them when I invoke the skill.
I set up the end result: a "good enough" article from the interview content. The point was to get out of my own way and let the LLM not only ask me questions, but make sure it had enough content from my answers to create that "good enough" content. It took away my excuse for not finishing a piece.
I made sure it had rules for how to conduct the interview. Primary rule: ask one question at a time until it has enough information for the content piece we are working on. I found that just telling Claude to ask questions until it had enough information would get me a wall of questions immediately. I know I tend to give overly complete answers and I can't do that when there are too many points to cover all at once. So I tried "one or two" questions and still found two to be too many.
I used to do one thing at a time, but now I've cut back to half a thing at a time. --GB
The other critical instruction was to only use my own words. If it generates anything like intros, headings, or transitions, it's supposed to mark them so I can circle back and clearly see what I need to address before hitting "go" to publish. No AI slop. Just light editing of typos and grammar, rearrange the chunks to flow better, and ask for more content or mark it as a sample of what might work there for me to work through myself.
The Reinforcement Loop
One of the unforeseen benefits of this flow was synergy with my own quirks. In this case, it was my tendency to give overly complete answers with my experiences of not giving the LLM enough context. I didn't want to give the LLM an incomplete idea and have it run off on some other line of thinking before I could finish that thought "properly." Instead I would write an essentially complete article in response to each question. It wasn't on purpose, it just worked out nicely.
I knew I wanted to give the LLM editor solid prose to work with, so I made sure to start with standalone sentences and incorporate the prompting question. But unlike my previous attempts at writing final copy, I would just let myself write because it would smooth any rough edges. And, ironically, there were very few rough edges for it to smooth because of all that permission to write a rough draft.
The first time I used the original interview style approach years ago, I was amazed at what came from it. It reminded me of an interview from a magazine and I realized it could be published that way. This was a slightly different beast -- this was me writing a nearly complete article in response to each prompting question. And because it was tied to the project files for the site, it meant the agent could just create a new file for that article and incorporate it into the site automagically. I was quite happy to have it generate all the metadata like titles and descriptions and I just clean them up as I see fit.
The irony of prompting the LLM to prompt me is not lost on me.
First Results and Scaling
Immediately after creating the Content Interview skill, I put it to use. I cleaned up a brain dump and got a resource page and anchor article about NLP for my main site. Then I cleaned up half a dozen more model pages I had already created without the skill.
My first new piece fully created with the new skill was on "learning loops." Then reworking the welcome letter and values pages. I managed 2000 words of near final copy in downtime throughout that first day -- and most days since! I'm making good use of the "golden hours" and otherwise wasted moments: while waking up, falling asleep, between meetings, in the evenings when my brain is winding down and I'm rewatching some stand-up or improv comedy.
Handing over the "good enough" decision to the LLM was emotionally challenging, but minimal risk. If I totally disagreed with how an article came together, I just didn't hit publish and all that was lost was the time spent futzing with it -- manually or continuing the conversation with my agent.
But the potential upside was massive. The thing that kept me from having more of an audience champing at the bit for more content is having more content out there for them to resonate with.
And that last phrase is doing the heavy lifting: content they can "resonate with." I had the technology to crank out tons of content that might also have been "good enough" to publish, but it has very little of me in it, so anything folks were resonating with was probably just noise from the AI, not my signal and myself.
With the risks mitigated and the benefits clear, I dove in. I started creating content and dialing in the process and the guardrails for the LLM. I knocked out several pages of rough content and tested different models. It took a few iterations to make sure the process was where it needed to be, but now I just run it.
Since the LLM was making very few real "decisions" and not really generating much prose -- just editing mine -- I didn't need any of the "frontier models" with all their capabilities and associated price tag. I could use smaller, cheaper, less capable models. Yes, I had to get the rules clear and make sure it was keeping my words (not rewriting it all into its bland tone), but I got it dialed in on a model small enough it can be run offline and locally.
Pennies per article in editing support and cranking out new pages as fast as I can get the thoughts out of my head. And it truly is my thoughts, not AI slop like I had been getting before that was only vaguely in the same direction as my thinking.
My big concern that the LLM's "good enough" would be closer to AI slop was not an issue. Instead, the process uses my own tendencies and my own words to hit a "good enough" that I didn't disagree with.
Could the output be better? Sure. Is it worth polishing prose that is untested rather than getting the concepts out there and getting feedback from real readers? I doubt it. In this case, I choose "done" and readable over "world class" and sitting in a file somewhere hidden from view.
This is what I mean when I say "Intelligence Augmented." Not handing over the important parts to the AI, but using it to support the critical work. Handing off the parts that we aren't good at or have too many quirks around while keeping ourselves in the loop to make sure it is actually doing what we want.