AI Guides

Reverse Prompting: How to Ask an AI to Write Prompts for Itself (and Other Models)

After ChatGPT Images 2.0 nailed the VRS logo, I asked it to write prompts for every other model. The text-only model that can't draw beat the image models that can. Here's how reverse prompting works.

2026-04-22 · 7 min read

After ChatGPT Images 2.0 created the VRS logo, I turned to it and asked: "Can you write a prompt that would create similar results in other models?" That simple question sparked the idea of reverse prompting — a meta-skill that lets a capable model analyse what worked and translate it into model-specific prompts for the rest of the ecosystem.

What is Reverse Prompting?

Instead of spending hours crafting prompts manually for each image generator, you ask a strong multimodal model (like ChatGPT) to examine a successful output, dissect the elements that made it work, and then produce tailored prompts for other systems. The result? A set of prompts that are already optimised for each model's syntax, strengths, and quirks.

How I Did It

Using the VRS logo as my test case, ChatGPT Images 2.0 generated a version that hit the brief perfectly. I then prompted it to reverse-engineer that success. The model delivered four distinct prompts:

Each prompt was adapted to the target model's strengths and limitations, proving that a single high-quality output can be reverse-engineered into a cross-model recipe.

The Four Prompt Styles

Here's what ChatGPT produced for each model, and why each one is different:

T1/T3: Flux Prompt (keyword-rich, detailed)

Front-facing symmetrical human brain made of futuristic circuitry and glowing neural pathways, transparent background. Organic brain silhouette constructed from metallic circuit traces, glassy conduits, and hundreds of luminous nodes. Nodes subtly blinking and pulsing with light. Main colour theme #534AB7 electric violet with cool blue highlights. In the exact centre is a realistic metallic CPU microchip, brushed steel surface, beveled frame, visible screws in corners, seamlessly embedded into the circuitry. Precision engraved text on chip: VRS large top line, COMPUTING smaller underneath, perfectly centred, etched into metal, clean spacing, not glowing. Circuit lines connect into chip from every side. Hyper detailed, premium sci-fi branding aesthetic, ultra sharp focus, crisp reflections, subtle bloom, centred composition, transparent PNG, isolated object.

Negative: background, shadow, extra objects, blurry, watermark, low detail, asymmetrical, floating text, glowing letters, distorted text, duplicate chip, messy circuits, cartoon

Why this works for Flux: Flux thrives on dense, keyword-rich descriptions. It processes the entire prompt as a semantic map, so more detail = more control. The negative prompt section is critical because Flux doesn't have a built-in negative prompt parameter — including it in the text helps steer generation away from common failure modes.

T2: SD 1.5 Prompt (weighted tokens, shorter)

(masterpiece), (best quality), ultra detailed, centred composition, front facing symmetrical brain made of circuitry, futuristic electronic brain, glowing nodes, metallic circuit pathways, transparent background, isolated object, electric violet colour theme #534AB7, blue highlights, hundreds of luminous nodes, subtle blinking light effect, realistic metallic CPU chip in centre, brushed steel texture, beveled edges, screws in corners, integrated into circuits, engraved text "VRS" on top, engraved text "COMPUTING" below, perfectly centred text, etched metal letters, sharp reflections, sci-fi render, clean premium design, sharp focus

Negative: worst quality, low quality, blurry, text errors, distorted letters, glowing text, floating text, duplicate objects, background, shadow, messy composition, asymmetrical, cropped, watermark, cartoon, noise

Why this works for SD 1.5: Stable Diffusion uses a different tokenisation system that benefits from parenthetical weighting — (masterpiece) gets more attention than plain masterpiece. Shorter is better here because SD 1.5 has a harder time with long prompts. I also heavily prioritise the negative prompt because SD 1.5 is prone to generating extra limbs, duplicate objects, and text artefacts.

T4/T5: Claude Prompt (structured intent + hierarchy)

Create a front-facing, perfectly symmetrical futuristic brain logo on a transparent background.

The brain should be made entirely of:
- glowing circuit pathways
- metallic conductive traces
- glass-like neural channels
- hundreds of illuminated nodes

The overall silhouette must clearly resemble a human brain from the front.

Colour palette: Use primarily #534AB7 (rich violet) with subtle blue electric highlights.

Centrepiece: At the exact centre of the brain, place a realistic metallic CPU microchip integrated naturally into the circuitry.

CPU details:
- brushed steel texture
- beveled edges
- precision screws in the corners
- premium industrial finish
- realistic reflections

Engraved text on CPU:
- Top line: VRS
- Bottom line: COMPUTING
- VRS larger than COMPUTING
- COMPUTING smaller and aligned beneath
- both words perfectly centred
- etched into metal with precision machining
- no glowing text, no raised lettering
- subtle engraved shadows only

Lighting: Nodes should appear subtly blinking or pulsing with light.

Style: Hyper-detailed, premium sci-fi branding, ultra sharp, polished, elegant, realistic materials.

Output: transparent PNG, centred composition, no background, no extra objects, no distortion, no asymmetry

Why this works for Claude: Claude processes information hierarchically. Give it structure — sections, bullet points, constraints stacked in priority order — and it produces better results than keyword vomit. Notice how each element has its own section with explicit constraints. Claude needs to understand what you want, not just pattern-match tokens.

T6: Nemotron Prompt (technical + concise + direct)

Generate a transparent PNG of a centred futuristic brain made from electronic circuits.

Specifications:
- front-facing symmetrical human brain shape
- circuitry forms entire brain structure
- glowing nodes across pathways
- metallic conductive lines
- colour palette: #534AB7 with blue neon highlights
- hundreds of bright nodes with subtle pulse effect

Centre object:
- realistic metallic CPU chip
- brushed steel
- beveled frame
- screws in corners
- embedded into circuitry

Text engraved on chip:
- VRS
- COMPUTING

Formatting:
- VRS large
- COMPUTING smaller below
- both centred
- engraved into metal
- no glow
- sharp clean machining lines

Style:
- ultra detailed
- photoreal sci-fi render
- premium technology branding
- crisp reflections
- transparent background

Negative constraints:
- no extra objects
- no background
- no blur
- no malformed text
- no asymmetry
- no duplicate chip

Why this works for Nemotron: Nemotron responds better to technical specifications than creative writing. It's an engineering brain — give it specs, not poetry. The "Specifications / Centre object / Text / Style / Negative constraints" structure mirrors how Nemotron processes information: structured data in, structured output out. This prompt scored 8/10 when fed to Flux, outperforming every other approach except the iterative ChatGPT method.

Benchmark Results: The Honest Take

After dozens of generations across Flux.1-dev (GGUF Q8_0), SDXL, Stable Diffusion 1.5, Claude, and Nemotron, the results are in — and they're nuanced.

Reverse prompting works. The prompts ChatGPT Images 2.0 generated for each model produced noticeably better outputs than anything I wrote manually. Composition improved, colour fidelity tightened, and the VRS-branded CPU chip appeared more consistently. On my local hardware, with the right tools and some patience (it's free, remember), I got genuinely usable results.

But ChatGPT Images v2 remains unbeatable for fine-tuning.

OpenAI's latest image model has a decisive advantage: iterative refinement. With a few simple conversational prompts — "make the chip more metallic", "sharpen the VRS text", "add a slight bevel to the chip frame" — I arrived at my target image in under eight attempts. The same level of precision on local models required 50+ seeds across multiple generation runs, prompt rewrites, and post-processing.

The gap isn't in raw generation quality. It's in the feedback loop. When you can see a result, describe what needs changing, and get a refined version seconds later, you converge on the target exponentially faster. Local models require you to regenerate from scratch each time — there's no "tweak this detail" on a 60-second Flux generation.

My verdict: Use reverse prompting to get your local models 80-90% of the way there — it's free, private, and getting better all the time. Then, when precision matters, bring in ChatGPT Images v2 for that final 10%. It's not cheating — it's using the right tool for the right job.

Coming Soon: The Reverse Prompting Tool

I am building a simple utility where you drop in a reference photo, and the tool spits out reproduction prompts for all major models — no manual tweaking required. Think of it as a prompt translator for visual generative AI. Stay tuned!


Found this useful? 👉 Follow @Raf_VRS for more AI Guides updates
👉 Support the work: ko-fi.com/rafvrs

Stop Scrolling. Start Building. #LocalAI #ReversePrompting #AIImages