DEV Community: shi warren

AI image generation is easy. Building the handoff layer is harder.

shi warren — Wed, 17 Jun 2026 12:05:35 +0000

AI image generation is easy. Building the handoff layer is harder.

A few days ago, I launched and continued testing a small AI tattoo idea tool.

At first, I thought the main challenge would be image quality.

Can the model generate something that looks good?

Can it follow the selected style?

Can it produce a clean result?

Those things matter, but after testing more real workflows, I realized they are only part of the product.

The harder problem is this:

An AI image is not automatically useful just because it looks interesting.

For this project, I did not want the tool to simply generate an image and stop there.

A tattoo image can be useful in different ways. Some users may save it as a strong direction. Some may bring it to an artist. Some may use it only to compare styles, placement, or lettering ideas.

So the product question became less about:

“Is this the final design?”

And more about:

“What does this output help the user decide next?”

That distinction changed how I looked at almost every part of the product.

The image is not the end of the workflow

A user does not just need an image.

They need something they can bring into a decision.

For example:

What is the core idea?
What style direction does it suggest?
Is the design too detailed for the intended size?
Does the lettering stay readable?
What should be adjusted before it becomes a real tattoo?

A generated image alone does not always answer those questions.

So I added a small “copy brief” action after generation.

Instead of only downloading or sharing the image, users can copy a short artist-facing brief. It includes the original idea, the selected style direction, the detail level, the composition direction, and notes about what may need to be refined.

The important part is that this brief is not an AI prompt.

It is not a long disclaimer either.

It is meant to be a handoff note.

That felt like a small feature, but it made the tool feel much closer to a real workflow.

Interface language changes the mental model

Another thing I noticed was that the wording of the interface changed how people used the tool.

For the tattoo lettering tool, the first version had one mixed input field. A user might type the exact words they wanted, but also add instructions like “with soft shading” or “no extra words”.

That sounds simple, but for image models it creates ambiguity.

Which part is the text to render?

Which part is only visual direction?

So I split the lettering workflow into two fields:

Exact Text — the words that should appear in the lettering
Supporting Details — optional visual direction like shading, spacing, ornaments, or underline flourishes

That small change made the product clearer.

The user is no longer writing a generic AI prompt. They are deciding what text they might use, and then adding visual guidance around it.

It also made the internal prompt safer. The model now gets a clearer instruction: render the exact text once, and treat supporting details as visual direction, not words to draw.

This is something I keep running into with AI tools.

The internal model language is often not the same as the user’s workflow language.

Lettering has different failure modes than general image generation

Tattoo lettering made this especially obvious.

For a general tattoo concept, a little variation can be acceptable. If someone asks for a wolf and moon reference, the exact curve of the moon or the exact shape of the fur may vary, and the image can still be useful as a direction.

Lettering is different.

If the user enters a name, date, initials, Roman numerals, or a short quote, the text has to stay readable. It also has to avoid adding extra words.

One test with a short numeric input showed this clearly.

The model followed the instruction to render only the exact text, but it repeated the same text multiple times in a stacked layout.

Technically, it did not add unrelated words.

But it was still wrong for the product.

That led to another prompt constraint: render the exact text once as one lettering design. Do not repeat, tile, stack, or turn it into a typography sample sheet.

This was a useful reminder that constraints need to match the actual user expectation, not just the literal prompt.

“Only draw this text” is not the same as “draw this text once as a usable lettering direction.”

Not every user needs AI generation immediately

Another useful lesson came from the free tattoo font preview page.

Some users are not ready to generate a custom AI lettering image. They just want to type a name, date, initials, or a short quote and compare how it might feel in different tattoo-style fonts.

That is a different workflow.

So I kept the Tattoo Font Generator separate from the AI lettering tool.

The free page is now more of an instant preview and export tool: users can compare font directions, copy the text, and download a simple PNG preview.

The AI lettering tool is for the next step, when someone wants custom composition, shading, ornaments, layout, or supporting details.

That separation matters.

If every page becomes an AI generator, the product becomes harder to understand. Sometimes the better entry point is a simple browser-side preview tool that helps the user make an earlier decision.

For this project, the free font preview is not a weaker version of the AI lettering generator.

It is the previous step in the workflow.

Model choice can change the product category

One of the most useful issues came from a real user prompt.

They tried an idea similar to:

Albanian eagle with a background of national hero

The fine line version looked closer to a clean tattoo direction.

The realism version, however, drifted into a dark poster-like image with a cinematic background.

The prompt already asked for a standalone tattoo-style image on a plain white background, but the model still leaned toward a realistic illustration scene.

That was a useful failure.

For this product, “realism” does not mean “photorealistic poster”.

It means a realistic tattoo-style direction that still works as something a person can evaluate, save, and bring into the next step.

I changed the realism route away from the model that produced more cinematic images and toward one that produced cleaner, more design-like outputs. The results became less dramatic, but much closer to the actual product goal.

That tradeoff was worth it.

A better-looking image is not always the better product output.

Constraints are part of the product

With AI image tools, it is tempting to think more freedom is always better.

But for this use case, constraints are what make the output useful.

The tool should avoid:

body mockups
skin previews
dark poster backgrounds
accidental extra symbols
misleading final-stencil claims
exact copies of copyrighted characters

Those constraints make the output less flashy, but more useful for the intended workflow.

A tattoo-style image can help someone explore a direction, but the final tattoo still needs human judgment around placement, size, readability, and technique.

This also made me add clearer wording around pop culture and copyrighted characters. If someone enters an IP-based idea, the tool should create an original tattoo-inspired direction, not an exact copy of a protected character.

That boundary matters.

The product is not just generation

The biggest takeaway so far:

The hard part of an AI image product is not always generating the image.

The hard part is deciding what the image is supposed to help with next.

For this project, the useful path is something like:

Idea → visual direction → comparison → handoff → refinement

That means the product needs more than a generate button.

It needs context, boundaries, readability guidance, style framing, export actions, and a better handoff layer.

I am still early, but the launch feedback changed what I am optimizing for.

Not just better images.

Better decisions after the image.

I recently launched this as a small project called AIMakeTattoo.

The project is here if anyone is curious: https://aimaketattoo.com

I’d be interested to hear how other people think about the handoff layer in AI tools, especially when the output needs to become part of a real-world workflow before it is used.

Google Indexed My Pages. Nobody Found Them.

shi warren — Tue, 09 Jun 2026 09:16:12 +0000

A few weeks ago I thought I had a technical SEO problem.

Pages weren't showing up.

Search Console looked empty.

So I spent a lot of time worrying about indexing.

Then the pages finally got indexed.

Nothing happened.

That was the moment I realized indexing and understanding are completely different problems.

I'm building a small browser-side utility.

The product removes metadata from photos and documents.

From a technical perspective the site was fine.

Pages were crawlable.
Sitemap was submitted.
Search Console showed indexing progress.

But search traffic still barely moved.

At first I thought Google needed more time.

Then I started looking at the queries that actually appeared.

Something interesting showed up.

Users searched for things like:

"remove location from photo"

"remove gps from image"

"remove author from pdf"

Almost nobody searched for:

"metadata processing"

"metadata extraction"

"browser-side metadata cleanup"

Those were the phrases I had been using when describing the product.

The product and the user were talking about the same thing.

Just in completely different languages.

That changed how I think about SEO.

I used to assume:

Indexing
→ Ranking
→ Traffic

Now it feels more like:

Indexing
→ Understanding
→ Ranking
→ Traffic

And the understanding step is surprisingly slow.

Especially for small sites.

The weird part is that AI makes this easier to miss.

Building is cheaper than ever.

Adding another page takes minutes.

Adding another feature takes hours.

Adding another tool feels almost free.

So it's easy to create more things.

But search engines still need to understand what those things are.

Lately I've been spending less time building pages and more time trying to understand how users describe the problem in their own words.

That has probably taught me more than any SEO guide I've read so far.

AI makes building faster, but semantic distribution is now the hard part

shi warren — Thu, 21 May 2026 04:49:33 +0000

Shipping products feels dramatically easier now.

Between AI coding tools, templates, and browser-side tooling, I can build and deploy things faster than ever.

But distribution feels harder.

Not technically harder.

Semantically harder.

A lot of the work now is:

understanding how users actually search
matching workflows instead of technical terms
figuring out where intent really exists
building trust before distribution channels suppress you

Building became compressed.

Attention became fragmented.

Honestly feels like distribution is becoming the real product skill now.

Most people search “remove location from photo”, not “EXIF”

shi warren — Fri, 15 May 2026 08:49:20 +0000

Most people search “remove location from photo”, not “EXIF”

While building a metadata removal workflow, I realized something interesting:

Most people never search for:

EXIF
IPTC
document properties
metadata

They search for:

“remove location from photo”
“hidden info in PDF”
“remove GPS from image”
“does this file still contain private data?”

The technical layer and the user intent layer are often completely different.

Developers think in systems.

Users think in outcomes.

That changed how I started thinking about both SEO and product design.

A lot of users are not trying to “inspect metadata”.

They’re usually just at the moment before sharing something publicly and wondering:

“wait… is there still something hidden in this file?”

Your photos may still contain GPS metadata after editing

shi warren — Sat, 09 May 2026 17:09:15 +0000

A lot of people assume cropping, resizing, or screenshotting a photo removes hidden information.

That’s often not true.

Many images still contain metadata like:

GPS location
device information
timestamps
camera details
EXIF data

In some cases that can unintentionally expose:

where a photo was taken
what device was used
when it was captured

I started noticing this more while working with:

screenshots
AI-generated images
product demos
travel photos
shared work files

One thing that surprised me is how many people don’t realize metadata exists until they accidentally leak something.

Privacy tooling on the internet still feels mostly reactive instead of preventative.

I think lightweight browser-side privacy workflows will become much more important as AI-generated and shared content keeps increasing.