The Sorcerer's Apprentice

For a 14-minute long analysis of this post by Google’s Notebook LM deep-dive AIs tap on the audio player below

---

Grudgingly, I’ve realized that Claude from Anthropic understands Midjourney better than I do. It’s taken considerable effort to surrender parts of my creative self, which was what I thought made me unique as an image creator, to an AI tool. 

How can I add value? More importantly, how can I inject “me” into an image if I relinquish the entire creative process to my tools? Where’s the fun if my only contribution to the process is typing a prompt and pressing Enter?

As I refined my creative process, this paradox remained central to my journey, and I have resolved it by shifting from a tool-centric view to a relational perspective—one that focuses on the relationships between the various AI collaborators in my creative workflow and me, the creative director.

---

Step 1 – Getting 3rd Party Feedback

Broadly, generative AI tools can be broken down into two categories: general-purpose and domain-specific. 

Claude from Anthropic and ChatGPT from OpenAI are built on generic large language models (LLMs). They are trained on everything. How to bake banana bread. Where does philosopher A’s insights overlap with philosopher B’s? Who wrote the song Stairway to Heaven? 

Midjourney, on the other hand, is domain-based and relies on a Large Image Model (LIM). It doesn’t know about baking, philosophy, or Led Zeppelin, but it does know, among millions of other imaging elements, how to apply Rembrandt’s style to a picture of a kitten wearing sunglasses.

The trick then is to integrate generic text-based AIs such as Claude with domain-specific AIs such as Midjourney.

The first step is to take a proven prompt, render it with the latest version of the Midjourney LIM, then analyze the result using Claude or your preferred generic AI.

Here’s the starter prompt I passed on to Midjourney V8.1.

Create an encaustic 3d bas relief post-impressionist painting of a beautiful woman with long flowing curly black hair, turning back to look over her shoulder at the camera, wearing a stunning, white wolf-inspired gown with delicate, flowing design and intricate details resembling the fur and structure of a wolf, long, curly flowing hair cascading down her back, catching the light and adding a luminous quality to the image, background is soft and blurred, ethereal, V@n G0gh, hdr A realistic picture of a ballerina wearing a chiffon dress made from dandelions with more details and holding an umbrella made from dandelions and dancing under a rain of dandelions in a romantic realistic style with a background.

---

Step 2 – Discussion

Claude was not particularly impressed with my long-time favourite prompt and responded like an ER nurse tending a newly arrived trauma patient. “Crash Cart!! Booth 6. Stat!!” Claude took action, revised the prompt with surgical precision, and made nine specific laser-focused changes.

• Unified the two subjects into one figure — same woman, same scene. V8 struggles with two separate conceptual subjects competing for dominance.
• Removed “HDR” and “realistic” — these directly fight Van Gogh/painterly. You can’t have both; the impasto texture wins anyway.
• Dropped “V@n G0gh” obfuscation — V8 doesn’t need it, and garbled spelling can confuse the model.
• Cut the duplicate hair description — appeared twice verbatim in the original.
• Replaced “camera” with “viewer” — more natural for painted scenes.
• Removed “more details” — weak instruction that V8 ignores; specificity does the work instead.
• Added “impasto brushwork” and “warm luminous glow” — these are the actual Van Gogh texture signals V8 responds to.
• Kept “encaustic” — it adds an interesting waxy depth that complements impasto without contradicting it.
• Added –ar 3:2 — consistent with your cinematic gallery work, and natural for a standing figure.
• The one remaining tension: the wolf gown and the dandelion ballerina are strong, competing aesthetics. If the render feels muddled, separating them into two images would actually let each breathe properly.

Accepting these insights, I continued and asked Claude to refine the prompt.

Claude’s refactored prompt:

Encaustic post-impressionist painting, Van Gogh style, a beautiful ballerina with long flowing curly black hair glancing back over her shoulder at the viewer, wearing a luminous white wolf-inspired gown with chiffon layers and intricate fur-textured details catching the light, dancing gracefully while holding a delicate umbrella woven from dandelions, golden dandelion seeds drifting down around her like snow, swirling impasto brushwork, warm luminous glow, soft ethereal bokeh background –ar 3:2 –v 8

I sent Claude’s revised prompt to Midjourney V8.1 and, after numerous AI hallucinations (Why is her head on backwards? Did I miss a meeting? “) and variations, I finally generated three decent images.

Claude responded to the results with a few follow-up questions, showing interest but not complete satisfaction.

• Do you want to push the wolf reference more explicitly (a wolf fur collar, amber eyes, that kind of anchor), or is the current feathery-gown interpretation the right landing place?
• The “over the shoulder” pose from the original prompt is partially present in image 3 — worth steering back toward that if the original intent still matters.

Both of Claude’s suggestions intrigued me, especially the emerging wolf references. 

However, I noticed something even deeper – motion. The flow of the woman’s dress and the dandelions now suggested the need for more kinetic energy in the image.

My reply to Claude: 

Let’s push the wolf reference and yes steer back to the “over the shoulder” perspective. Also, increase the kinetic.

Here’s Claude’s revised prompt, along with a recommendation to use a style reference (Midjourney’s —sref parameter) linked to the second image in the gallery of three shown previously:

Encaustic post-impressionist painting, Van Gogh style, a beautiful woman mid-spin, back to the viewer, twisting sharply to look back over her shoulder with a direct gaze, long curly black hair streaming wildly in motion, wearing a white wolf-inspired gown with a dramatic silver-grey wolf fur collar and flowing wolf pelt cape, amber wolf eyes, intricate fur-textured chiffon layers billowing in a vortex, golden dandelion seeds spiraling violently around her like a storm, turbulent swirling impasto brushwork, churning cobalt blue and gold background whorls, encaustic wax surface texture, warm luminous glow –ar 3:2 –v 8

Which produced this ‘getting-close-to-perfect’ image.

Both Claude and I appreciated the result, though I found the blurry face distracting. 

My feedback to Claude:

In answer to your question, this image is 95%, but her face isn’t as clear as I’d like it to be. If we could improve the clarity of her face using a remix prompt seeded with this image, we might be able to fix that.

---

Step 3 – What Was Old Made New Again

Claude thought about my comment for a few seconds and created a third prompt that included the image weighting (–iw) option. Image weighting lets you specify how much the style reference guides the output on a 0–3 scale. I used 2 to emphasize the reference, instructing Midjourney strongly but not exclusively.

Claude’s revised prompt

Encaustic post-impressionist painting, Van Gogh style, a beautiful woman mid-spin, back to the viewer, twisting sharply to look back over her shoulder, sharp facial clarity, luminous skin, well-defined features, expressive dark eyes with amber highlights, soft parted lips, long curly black hair streaming wildly in motion, wearing a white wolf-inspired gown with a dramatic silver-grey wolf fur collar and flowing wolf pelt cape, golden dandelion seeds spiraling around her like a storm, turbulent swirling impasto brushwork, churning cobalt blue and gold background whorls, encaustic wax surface texture –iw 2 –ar 3:2 –v 8

The noticeable defect in this prompt, and Claude can be forgiven, considering how fresh Midjourney V8.1 is, was that the version 8.1 of the LIM, at the time of writing, does not support image weighting. Ack! Fortunately, even though I’ve been working exclusively with the Midjourney V8.1 website for over a month, my previous skills on Midjourney Discord, where Midjourney V7 still lives, remained sharp.

The results were amazing and better than what V8.1, for this prompt, was able to deliver.

I uploaded all four MJ V7 images to Claude and provided the following feedback:

What a difference a model makes. These four images were rendered with Midjourney V7. V8.1 does not yet support image weights (–iw). Many of the images V8 generated, I did not like (lots of hallucinations). However, the first four images rendered with V7 using –sref and –iw are amazing. No duds.

Claude was equally impressed with the images and delivered some pretty illuminating feedback on all four.

Image 1 — The face clarity goal is fully achieved here. The gossamer dandelion wisps are the most botanically accurate in the entire series — delicate filamentous seed pods, exactly right. The luminosity on her skin is exceptional.

Image 2 — The strongest synthesis of all your objectives. Face clear, over-the-shoulder gaze locked, wolf fur collar dominant and well-defined, impasto brushwork kinetic. This is the series centrepiece. (Ed: Not so fast, Sparky. You ain’t seen nothin’ yet.)

Image 3 — V7 made an autonomous creative decision and materialized an actual white wolf as a spirit companion. That’s remarkable and arguably the most narratively rich image of the entire series. The wolf’s blue eye, mirroring her gaze, is an unscripted gift.

Image 4 — the most physically kinetic pose of all four. Arms raised, direct confrontational gaze, the fur dissolving into the gold storm. Different energy — less dance, more defiance.

Claude’s comment about the third image resonated with me: the white wolf as a spirit companion for the woman. That felt like a breakthrough.

Claude also noticed something that led to this blog post—exploring the synergy of mixing Midjourney V8.1 and V7. This whole process illustrates my main argument: effective integration and collaboration with AI can expand creative possibilities and deepen my own role in the artistic journey, rather than diminishing it.

Claude’s take:

The workflow lesson here is significant: V7 with –sref + –iw 2 is currently the correct tool for seed-controlled iteration. V8.1’s lack of –iw support makes it better suited to fresh generations from strong prompts, while V7 owns the refinement/remix lane. Two-model workflow: V8.1 to discover, V7 to refine.

Image 3’s wolf companion opens an interesting question — is that a direction worth pursuing deliberately, or do you want to keep the wolf as clothing/metaphor rather than a literal presence?

I replied:

I’d like to explore the wolf companion further, as you noticed, from a spiritual companion perspective. I’ll stay rooted in V7 for the renders and still refer (–sref) to our source image created by V8. Best of both worlds. Please create a new prompt that delves deeper into her journey with her spirit wolf companion.

Claude crafts the final prompt:

Encaustic post-impressionist painting, Van Gogh style, a beautiful woman on a sacred journey, back to the viewer, turning to look over her shoulder with a knowing gaze, long curly black hair streaming in motion, wearing a white wolf-fur gown that dissolves seamlessly into the coat of a great spectral white wolf walking silently beside her, wolf and woman sharing the same luminous amber-blue eyes, their spirits visibly intertwined, the wolf’s form half-flesh half-golden light, dandelion seeds and impasto brushstrokes flowing between them like shared breath, neither leading nor following — bonded equals, churning cobalt blue and gold background whorls, encaustic wax surface texture, sacred luminous glow –ar 3:2 –v 7 –sref [your url]

The result, again, was four wonderful images with no hallucinations, which I uploaded to Claude for its analysis.

Claude was as thrilled as I was:

Four from four is extraordinary — this series has found its mythology. Image by image:

Image 1 — the most tender. Her eyes closed, face turned inward, the wolf’s amber gaze doing the watching for both of them. It reads as trust — she doesn’t need to look because the wolf sees. The luminous merge at her shoulder, where fur becomes gown, is the most seamless of the set.

Image 2 — the most regal. Her profile against the gold storm, hand resting on the wolf’s neck — not gripping, just present. The amber wolf eye is now a signature element of the series. The gown-to-wolf-fur dissolution is fully realized here.

Image 3 — the most cinematic. The wolf dominates the foreground; she recedes slightly — for the first time, the wolf leads. That power dynamic shift is narratively interesting and visually unique in the series.

Image 4 — the most intimate and my personal favourite of this batch. Face clarity is the best across the entire series in all renders. The wolf’s head tucked close to her shoulder, both facing the same direction — the “bonded equals, neither leading nor following” intent from the prompt is perfectly expressed. The dandelion seed clocks embedded in her hair are exquisite.

Where this series stands now: you have a complete, publishable gallery. The progression from solo dancer → wolf-gowned figure → spirit companion revealed tells a coherent mythological arc across twelve images.

TLDR:

• Having a text AI, Claude, analyze a Midjourney prompt reveals not only defects in your prompts but also new directions.
• For this series, Midjourney V8.1 was great at rendering the basic look, texture, and motion, but was overwhelmed by the scene’s complexity. For Midjourney V7, using a Midjourney V8.1 image as a reference, the real beauty was surfaced.
• I’m becoming a creative director leading a gaggle of AIs. Not a prompt engineer. A creative director.