There's a problem nobody talks about in the AI hype cycle: consistency.

Everyone talks about how they used AI to generate assets. These assets include billboard advertisements, social media images, and one-pagers. Nobody talks about the number of iterations it takes and how the output still looks completely different from yesterday, even though you used the exact same prompt.

I experienced this first hand multiple times trying to produce social media graphics for a client.

The Carousel of Frustration

We had a client-approved format locked. We had aligned on various elements, like customer logo placement, headshot position, quote layout, and brand colors. My job was to produce a series of these graphics efficiently using AI image generation tools to showcase how a variety of customers were getting value out of my client's product.

Simple enough, right?

Except tools like Nano Banana and Manus.ai would keep generating images which looked completely different or came up with their own format, completely ignoring the approved format. I would even feed them an example image and use the exact same prompt that I previously used. And every single run would come back slightly off. The output was all kinds of variations: different font type or size or color, logo drifting to the wrong corner, quote text misaligned, headshot too large, too small, cropped wrong, streaks at the bottom, incorrect logo, incorrect headshot. Every iteration had some issue or the other. This was taking up finite time and energy. As a solo marketer, every minute spent prompting AI tools to produce incorrect outputs builds frustration fast.

Every iteration burned tokens and time. What should have taken 20 minutes was stretching into hours, and I was no closer to a repeatable process.

The Shift: Ask the AI to Write the Prompt

At some point, the obvious hit me. I was spending all my energy trying to describe the output to the AI. What if I flipped it?

Instead of me writing prompts and hoping the tool interpreted them correctly, I went to the AI tool and said: look at this image, understand every structural decision in it, and write me a prompt that would reliably reproduce this layout.

The tool generated a prompt which was so detailed that, as a human, I would have never been able to come up with it. It worked.

That is reverse prompting. And it changed everything.

Here is the actual prompt I now use every time:

Reverse Prompt Template 
Recreate the attached testimonial graphic as a near-exact template
match to the reference image. Keep the same canvas size, background,
composition, padding, spacing, fonts, color palette, shadows, photo
treatment, logo placement, and overall styling.

Replace only the following:
- top-left company logo: {company website URL or direct logo file/URL}
- upper small quote text: {text or `n/a`}
- main quote: {main testimonial text}
- highlighted phrase inside main quote: {highlighted text}
- name: {full name}
- title/company: {job title, company}
- headshot from image URL or attached file: {URL or file name}
- bottom-right product logo: {direct logo file/URL, or `reuse existing from reference`}

Preserve these design specifications exactly:
- size: {exact width} x {exact height} px
- background: {color/gradient description if known, or `same as reference`}
- quotation mark placement/style: {describe or `same as reference`}
- intro text color: {hex or `same as reference`}
- main quote color: {hex or `same as reference`}
- highlight phrase color: {hex or `same as reference`}
- headshot treatment: {shape, border, size, or `same as reference`}
- name/title block placement: {same as reference}
- bottom-right product logo scale and margin: {same as reference}
- no layout changes, no style changes, no additional elements

Important requirements:
- Keep the final result visually indistinguishable from the reference
  template except for the swapped content.
- If any logo or headshot asset is missing, ask for it before proceeding.
- Reuse only assets explicitly approved by me.

This prompt did not come from me sitting down and writing an image spec from scratch. It came from asking an LLM to study a finished image and extract the rules behind it. Working with AI, I built a prompt specification, not just a one-liner, but a structured brief that captured the exact logic of the asset.

The AI codified the image, creating an instruction manual for itself.

What Happened Next

With the new prompt, I went from 15+ iterations to two or three. What used to take hours now takes about 10 to 15 minutes.

More importantly, the output is consistent. Not perfect-every-time consistent, because generative tools currently always have some variance. The output is close enough that my edits are minor, not fundamental. I just need to fine tune rather than rebuild or start from scratch.

Most people overlook AI prompts. They usually lack the details and precision needed for consistent and high quality output. There needs to be a framework. It is an investment that pays back on every subsequent asset in that format. If you produce it regularly for a client, the ROI is obvious.

The Broader Principle

This technique is not limited to images. It applies anywhere AI output needs to be consistent. If you have a LinkedIn post format that performs well, extract the structural rules so you can replicate the pattern without copying the content. Have an email template your team keeps unnecessarily reinventing? Codify the logic once and stop starting from scratch.

When you need AI to do something reliably, stop describing outputs and start documenting the specs. Create a framework for the system that produces them, and use it to rinse and repeat.

The Takeaway

If you are past three iterations trying to reproduce a format you already know works, the problem is not the tool. You are prompting by feel instead of by specification.

Use AI to help you write the spec so you get consistent output. Then use that spec every time.

Reverse prompting is not a workaround. It is a skill. Learn it, use it.

Questions or thoughts? I'd like to hear where you're stuck. amit@curioinsight.com