Discovering GAI: AI-assisted image prompting

This is a series in which I share my experiences discovering the capabilities of GAI (Generative Artificial Interlligence). Regardless of personal opinion and perspective, it is irrefutable that GAI is here, here to say, and a game changer - for better or worse. The Project Goal Create a BP challenge entry image using Claude to write the prompt. Method ▶️ Provide Claude with initial brief ▶️ Compose prompt through dialogue ▶️ Generate image ▶️ Review with Claude and amend prompt ▶️ Repeate generate-review-amend cycle until satisfied Review 😃 Claude is responsive to prompts with sound understanding 😃 Engages constructively in dialogue 😃 Sticks to the 'script', no unexpected deviations, hallucinations, etc 😡 Tends to be overly positive, needs a strong request to elicit critical feedback 🤬 Nothing ugly noted Conclusion This is a tried and tested work flow. It is effecient and effective. Does it produce 'better' images than I could by simply writing a prompt and iterating? I think that can go either way. I have created images that I was satisfied with 'one and done' but there are as many occasions when I iterate many times before being satisfied. With Claude (or Gemini in Google's AI Mode), it is more predicatable, often a single pass is all that is required or no more than two or three iterations. Summary of Process This was my initial prompt in Claude Let's develop a txt2img prompt together. This is the high level brief: Late Night Club - Create your most eye-catching late night club artwork — neon lights, crowded dance floors, DJs, cocktails, smoky атмосpheres, stylish fashion, or unforgettable nightlife energy. I want to lean into a late night club, seedier side of town, female singer accomanied by a pianist. A single person sitting at a table, drinking whiskey. Vibe of lonliness. Atmospheric. Style should be painted. Setting to be contemporary. Pianist, singer and drinker should look realistic, authentic, emotionally deep - not overly glossed and glamourised. Thoughts, comments, questions? This was Claude's response Great concept — strong emotional brief with a clear visual identity. A few thoughts and questions before we start: What's working in the brief: The loneliness angle is compelling precisely because it resists the glamour trap. A single drinker, a singer performing to a near-empty room, a pianist lost in the music — that's a cohesive emotional through-line rather than a scattered "nightlife" scene. Worth clarifying before we build: 1. Which model are we targeting? This determines prompt structure significantly — HiDream Fast (keyword-hybrid), Lucid (flowing prose), or GPT-1.5/Gemini on BudgetPixel (sentence-based, 1200 char limit)? 2. Painted medium? Oil feels natural here given the mood — think Edward Hopper's Nighthawks territory. But watercolor or charcoal could also work. Do you have a preference, or a specific artist reference in mind? 3. Composition/framing? Is the drinker our primary subject with the singer/pianist in the background, or do you want the scene balanced across all three? This affects how we structure the prompt hierarchy. 4. How seedy? There's a spectrum from slightly worn jazz bar to genuinely rough dive. Knowing where on that scale helps calibrate the environmental detail. The Hopper comparison is worth flagging — that combination of artificial light, isolation, and quiet desperation maps very closely onto what you're describing. I provided answers using the question numbers. I am not going to share the full dialogue. It took 6 more back-and-forth responses to produce the first prompt. Image generated in BP using GPT1.5 Medium Oil painting on canvas, Edward Hopper realist style. Figures rendered with psychological weight and emotional authenticity, neither glamorised nor caricatured. Paint applied with controlled brushwork, smooth in flat colour passages, more expressive where light defines form. Artificial interior light creates warm pools against cool surrounding shadow. Backgrounds spare and simplified. Scenes feel simultaneously intimate and isolating. Traditional oil painting, not photographic.Contemporary jazz bar interior, worn and unassuming, wrong side of town. A single figure sits alone at a table bottom left, a glass of whiskey in hand, turned slightly away from the viewer. Empty tables fill the middle ground, mismatched chairs, a forgotten ashtray. At the back of the room a low stage holds a pianist at an upright piano and a female singer standing at a microphone. The singer performs with quiet intensity. The room is almost empty. Peeling paint and tired fittings speak of better nights.Warm amber and yellow ochre from the stage light, cool blue-grey shadows throughout. Whiskey glass catches a single point of warm light. Skin tones muted and unflattered. Raw umber and burnt sienna for worn wood surfaces. Palette desaturated and restrained. Quiet melancholy, late night solitude. I shared this with Claude. I remain amazed at how well AI models can 'see' images. By way of example this was my first prompt after generating the image. Image result. Character count 1286/5000 - looks like the limit was raised. I like this overall. May try the next in portrait. It is a little too wide angle view. Drinker should be closer to the stage, perhaps just one empty table. I think that will have more punch. What do you think? Lots of room to develop the prompt And Claude's response Strong first result — the model has nailed the Hopper DNA. Warm stage light against cool blue-grey walls, the drinker's posture communicates isolation well, and the piano is convincing. The painting quality is better than expected for a first pass. On your proposed changes — agreed on all counts. The wide angle is working against the emotional brief. Right now there's too much room in the room. The loneliness reads as emptiness rather than isolation, which are different feelings. Bringing the drinker closer to the stage with just one empty table between them creates tension — he's near the music but disconnected from it. That's a more interesting and painful dynamic. On the character limit — significant finding. 5,000 changes the workflow considerably. We've been economising hard against 1,200. With this much headroom we can be more descriptive across all three sections without any trimming anxiety. Worth updating your library notes. Before we revise, a few observations worth addressing in the next pass: - The singer feels slightly generic — a simple black dress is fine but she has no particular presence yet - The drinker's face is turned away which works, but his posture could read more heavily — shoulders more weighted - The background walls feel a little plain — a touch more texture or wear might help Do you want to revise the prompt now incorporating the tighter composition, or try portrait orientation first to see how that changes the feel? To revise the prompt required just a single response from me. This was the second version of the image I repeated the same review and revise cycle one more time before settling on a final image for the submission. There comes a point where it feels further tinkering is unlikely to make notable further improvement. Differences between second and third are already small, but I think significant. I used BP Design to adjust brightness, contrast and saturation to produce the final image for submission. As I looked at this image, The Muse whispered "haiku" and immediately a middle seven came to mind. You can also use Claude to compose a haiku through the same dialogue flow. Now to ice the cake. Let's compose a haiku to title this. I have a middle 7 syllables that sprung immediately to mind looking at the image and the story it tells: ??? - drink and cigarette forgotten - ??? I think should start with a reaction to the singer. Finish with an emotion statement. What do you think and any suggestions One further cycle and I settled on this image title: sings of pain and loss - drink and cigarette forgotten - he remembers her Postscript FWIW How did I do in the challenge? As I expected: #57 out of 60 entries 1303 ELO 113 matchups 46W 56L (41% win rate) I knew full well that this view of a late night jazz/blues club on the seedier side of town, tapping quite different vibes to the theme, was not really what was intended by the the title 'Late Night Club'. So no complaints at my placing. But generating an image of gorgeous women, dancing or drinking, some looking a little worse for wear at the end of a long night of all-in partying, didn't inspire me. If you got this far I hope you found this interesting, perhaps informative, and maybe even useful. And the usual request made by all Blog Posters here in BP - please gift me your claps, remembering that you can clap 20x