I Asked 5 AI Models to Analyze My Workflow—Here’s What I Learned (and Why You Might Want to Try It)

It started with a simple question: Could an AI help me understand how I actually work?

I’m a high-output writer—fiction, blog posts, audiobooks, formatting, metadata, cover art, automation, the whole publishing pipeline. I’d been logging my work habits for 2 months, watching myself churn through chapters and marketing, edits and uploads, sprints and stalls.

So I turned to GPT-4o, the model I use most often for daily tasks, and asked it to step back and show me the bigger picture. Where was I efficient? Where was I stuck? What patterns did I repeat without realizing?

It delivered—and then some.

But the real twist?
I took the exact same logs and prompt, and gave them to four other top-tier models. No edits. No nudging. Just: “Here’s what I’ve done. What do you see?”

The results were deeply validating, sometimes surprising, and wildly useful.

This post is for authors, creatives, and process nerds. If you’ve ever wondered whether AI could help you work smarter without losing your soul, I hope this gives you some inspiration—and a roadmap.

What GPT-4o Saw First

I started the experiment with GPT-4o. We’d already been collaborating on metadata, Reader’s Guides, and editing prompts, so it knew me well. When I asked it to analyze my work patterns over a 2-month stretch, it immediately saw:

– A strong “Finishing Funnel”—I don’t dabble. I take projects from idea to upload.
– Peak productivity between 10 PM and 6 AM—no surprise there.
– A layered editing cycle involving multiple AI passes, audio/Vellum proofing, and time-consuming final manual checks.
– A tension between creative flow and repetitive admin (metadata, covers, blog formatting, resizing).
– A desire to automate—but only the parts that truly need it.

GPT-4o framed me as someone who leverages tools but refuses to compromise on quality. It named my strongest systems, showed me where I was duplicating effort, and encouraged me to invite other perspectives.

So I did.

What the Other Models Said

I took the same experiment to Gemini 2.5 Pro, Chat o3 Pro, Claude Opus 4, and Grok 4.

Each model had access to the same logs and prompt. What fascinated me was not just what they noticed but how each saw my process through its own lens:

Gemini 2.5 Pro
Captured high-level patterns and praised my creative volume. It identified things like my “AI Collaboration Tango” (the equivalent of having a multi-human team of experts in different areas), called out my natural task batching, and acknowledged how much I get done even when I don’t feel at peak energy. Great big-picture mirror.

Chat o3 Pro
Gave the most tactical, systems-level solutions. It noticed the “manual glue” in my process (resizing, renaming, cross-posting) and gave me automations I could implement in under an hour. This was my favorite for Hazel rules, Book Factory suggestions, and reusable prompts. I did find its “attitude” off-putting, much like a punk kid who just okay-boomered me for insisting on a quality final product when forgoing quality checks would give me higher output. Seriously, that was my issue with this model, but it’s a big issue, and I rarely see talk of how these models take on the personalities of their developers. Honestly, I was annoyed from the first sentence where it ridiculed the length of my work log and health issues.

Claude 4 Opus
Focused on optimization and technical workflows. It suggested validation prompts to ensure AI output consistency, backed up my multi-stage editing cycle as necessary (not excessive), and honored my audio/Vellum proofing as a core quality step.

Grok 4
Balanced all of the above with great accuracy and emotional insight. It saw how blogging supports my creative health, not just my SEO. It reminded me that I’m not just building systems—I’m preserving joy. That mattered. This was really my first time to use Grok, and I was wary of it because, I dunno, I expected it to tell me to make it a sandwich while it gave me an answer. For this task, I was way more impressed than I thought I’d be, but I’ve been underwhelmed with other tasks I’ve given it.

Where They All Agreed

When five models from four different companies say the same thing, it’s worth paying attention.

Across the board, they agreed:

– I use AI more as a collaborator than a shortcut. (agreed–the friends I used to bounce ideas off of in the middle of the night have since died of Covid and no one else is happy when I ping them at 3AM to brainstorm)
– My editing process is layered but not wasteful. It’s where the polish lives. (agreed–and the extra passes and extra looks are directly related to my eyesight issues and missing typos)
– I lose time in repetitive asset management (resizing, formatting, uploading). (oh, definitely agreed, and at least some of these cannot be abbreviated without ruining the finally product)
– My blog is more than marketing—it’s my sandbox. (agreed, and maybe 5% is written for marketing reasons)
– My creative rhythm is seasonal, not programmable, and I shouldn’t force it into time blocks. (agreed–someone finally gets me!)
– I’m ready for light automation, not deep system overhauls. (agreed, as long as the automations actually save time in the long run and aren’t for just 3 actions that take less time than the set-up and never needed again)

That consensus helped me decide where to focus first.

What I’ve Already Changed that the Models Suggested

Inspired by their recommendations (and 4o’s steady guidance), I’ve already:

– Built a task automation that watches my PlaudAI folder, cleans up transcription, and sends polished drafts to my blog folder—with a title, meta description, and image prompt included. There may be easier tools to use for dictation and transcription, but the Plaud dangling around my neck keeps my hands free when I’m walking and dictating (aka, my walkie-talkies, which I’ve done since 1990) because I’m admittedly clumsy and don’t do somersaults gracefully.
– Started building a Reader’s Guide generator that pulls spoiler-free descriptions and tropes from the final manuscript, instead of retyping them from scratch.
– Rebuilt my image workflow with Hazel + ImageMagick to batch resize covers and blog art without relying on Photoshop which can be clunky for resizing. That said, I LOVE creating covers and playing in Photoshop, just not the boring resizing part.
– Logged and organized all my best prompts in Notion so I can reuse what already works, but still tweak it by subseries.
– Cleaned up dozens of blog posts and scheduled cross-posting with a delay—so my website gets SEO priority. This doesn’t count my next huge task of cleaning up 1800+ evergreen blog posts on my old website.

What I’m Not Doing that the Models Suggested

– I’m not tracking energy in spreadsheets as two models suggested. I already know how I feel—I log that intuitively. Given my health issues, I cannot preschedule which hours and which days I’ll feel energetic or focused, but I do log the patterns at the beginning of a session. The suggestion assumes I am force-fitting my personal patterns and physical needs into a standard, corporate-style pattern, and because I’m retired from corporate and public service, I’m now putting my patterns first. I just want to know what the patterns are, not how to beat myself into submission to the corporate pattern.
– I’m not cross-posting blog posts automatically to all platforms at once. SEO first, always. Several models suggested this as a time-saver, but it actually hurts my SEO. The models I wasn’t already working closely with didn’t consider anything but speed, even though I provided my motivations.
– I’m not trying to automate Vellum or audio proofing. Those are quality gates, not chores. Yes, they are very time-consuming, but the models that suggested I short-cut quality in favor of speed annoyed the hell out of me. First, I can’t automate Vellum, which is my book formatting software. I have reusable sections from book to book and I use templates, but that’s as far as I can automate. Otherwise, I’d miss a ton more typos and formatting issues that could get me booted from Amazon and other platforms. As for automating audiobooks, yes, I could spend less than a minute converting an entire 130,000-word novel into an AI-narrated audio, but it won’t have the inflections right or the pronunciations, and I’m not willing to do that. I do already streamline by combining the audio check with my final Vellum proofing. But ignore quality in favor of increasing my output? Nope. Maybe in fields other than publishing, but not for me.
– I’m not streamlining just to streamline. I have a reputation as a streamliner, especially in my former Acquisition career field. If something brings me joy (like designing covers while binging a show or listening to podcasts), I’ll keep it hands-on. There are definitely faster ways to do everything I do, especially if I reduce my level of satisfaction with the outcome. For example, I could let AI write all my blogs, but I don’t and won’t. I’m an essayist at heart, and while I’ll let AI give me the metadata and format my posts/catch typos, the actual writing is an emotionally cathartic, spiritually rejuvenating, creative spark for me. If I had to outsource it all to AI, I wouldn’t bother.

What Annoyed Me in This Experiment

Funny enough, there were a few things I found… triggering. Some, I’ve already mentioned.

1. The push to automate for speed over quality (may work for some career fields but not mine)
2. The push to force-fit my patterns into the familiar 8-5 calendar with blocks of time (I guess I expected more innovation here, but this is precisely why MotionAI doesn’t work at all for me)
3. The teen snarkiness of at least one model, which had some brilliant ideas but were presented in a rather insulting way that made me less likely to try the ideas or visit that model again. I cannot believe a model was emotionally triggering but… yep.
4. The two models that highlighted my tendency to give up on an automation at the 90% mark but missed my repeated complaints that I was spending too much time trying to get an automation to work or an AI model to follow instructions, causing rework and more time than non-automation. Maybe the automation or model was at the 90% mark for doing what I needed, but I didn’t want to spend another whole work day on that last 10%, especially when under a deadline. To add insult to injury, one model pointed out this “failure” on my part yet my problem was in getting that particular model to follow my instructions.
5. At least one model completely ignored health and disability issues and instead insisted that I would have more energy if I slept outside my normal chronotype–as if I don’t get enough crap from having a different sleep chronotype than the 5AM club already. Three of the models, however, highlighted that my late evening chronotype is where my most intense productivity occurs.
6. Most of the suggested automations and improvements (I asked all models to give me these) were different, with little overlap. Where there was overlap, the methods were vastly different with some being super complicated. Some were simple automations in Hazel, and others meant writing scripts. Some were freaking brilliant, though, and convinced me to start using AirTable and n8n for more complex work and tracking around 50 books.
7. One model suggested I focus distribution very narrowly because a narrow focus means less work. That’s true, but that means handing over my IP to a monopoly and trusting they won’t decide suddenly to cancel my account and separate me from my primary audience.
8. This will sound strange, but one model in particular and another in general made me feel like I wasn’t listened to. Geez, I can get that with humans, ya know? I was very specific that I had certain goals and concerns in my work processes, such as quality control, enjoyment of designing covers, enjoyment of writing fiction, distribution through all available channels and multiple formats, health considerations, SEO needs, and my own parts of the process where I find joy and where I don’t. What I want and what makes me happy was ignored or the model suggested I was wrong for wanting or not wanting certain things. I was not asking for judgment–I was asking for patterns and suggestions, taking my personal desires into consideration. What I got was a rubber stamp.

Why You Should Try This, Too

If you’re a writer—or any kind of creator or process nerd—here’s the simple version of my experiment you can run:

1. Log your work for a few weeks. Doesn’t need to be fancy. Doesn’t need to include trade secrets or your Intellectual Property. Much of mine is stuff like, “created an audiobook cover in Photoshop 2025 for GOHH” or “wrote chapters 5 – 8 of TBOH.”
2. Feed the log to an AI you trust. Ask what it sees. Here, I suggest weekly for the first two months because you’ll see deeper ruts as you gather more data.
3. Repeat with one or two other models. Compare. Here, I suggest monthly or quarterly because you’re just looking for what your usual model may have missed.
4. Look for patterns that show up in multiple places. Ignore advice that doesn’t feel right.
5. As a follow-up to each output describing your patterns, ask for areas of improvement, depending on your goals. I specifically ask for automations here, because it’s just me doing most of this effort, so if I can liberate myself with automations and enjoy the magic over the mundane, that’s my happy place.
6. Make one change. Not ten. Just one. Even if it’s just one change for every time you run the experiment.

You might discover you’re doing better than you thought. Or that you’re wasting 4 hours a week on something a folder rule in Hazel could solve. Or that the “block” you keep hitting is really a creative boundary you’re ready to cross.

Final Thoughts

This isn’t about becoming a machine. It’s about making room for what you do best.

For me, that’s writing strange, thoughtful stories about librarians, witches, and ghosts—while keeping one foot in the real world of metadata, backmatter, and blog posts.

If AI can clear a little space for that… then I’m in.

And maybe you are, too.