What Do We Use AI For?

For repetition, scanning and what slips past. Not for the decision or the final word.

Owning a tool doesn't mean using it for everything.

We use AI. We are writing that as a plain statement, not a marketing headline, because whoever is going to live in a house has a right to know who built it.

For us AI is not an architect; it is a well-sharpened hand tool. We make the decisions, we do the work. It takes on the repetition, the scanning, and the things that slip past.

We put it to work in three areas.

01
Code & Development
We don't have AI build the site. We build the structure.
Refactoring, tests, debugging and security scanning: anywhere that asks for repetition and rigour.
What escapes a human eye usually escapes it out of fatigue. We close that gap.
In short
The architecture is ours. AI is on for refactoring, tests, debugging and security scans, to catch what slips past.
02
Research & Testing
We test a decision before we defend it.
We use it to try user flows, run accessibility and performance audits, and compare alternatives quickly. The goal is not to produce more options; it is to rule out the wrong one early.
Put plainly
Flow trials, accessibility and performance audits. Not more options, earlier elimination.
03
Visuals & Video
Speed shouldn't steal from the face of the work.
We draw on AI for brand visuals, short videos and motion pieces.
But every frame passes through our hands: the direction, the rhythm and the final word are ours.
In one sentence
Support on brand visuals, short video and motion work. Direction, rhythm and the final word stay ours.

The tools we use

There is no transparency without naming names.

Tool · use · data policy

Tool	What for	Is your data used for training?
Claude	Code and research	No
Higgsfield	Visuals and video	Yes, on the standard plan
Raylight	Interface videos	No

We write the third column too. Higgsfield's standard plan does open data to model training; what that means for the way we work is in the data section below.

This list is not fixed. When we drop a tool or try a new one, we update it here; the date at the foot of the page tells you when that last happened.

The maintenance behind this page.

A number is more convincing than a sentence.

weekly maintenance runs: 16
security advisories closed: 21
packages updated: 77

From this site's own maintenance records, 27 April – 2 September 2026. A record, not a benchmark: the sum of the weekly runs.

This is area 01 above in concrete form. Dependency scanning is work that rewards repetition, which is exactly why AI does it here: the distinction this page describes, applied to ourselves.

What you hand over, and what we hand back, stays yours.

Your data

The three tools are not the same, so we write down what each one does.

Everything you give us (text, visuals, code, brand files) stays yours.

Claude, which we use for code and research, and Raylight, which we use for interface videos, do not open data to model training. Higgsfield, which we use for visuals and video, does on its standard plan. We would rather write that down than leave it out.

So your brand files, your code and your unreleased work are never uploaded to Higgsfield. What gets made there is raw imagery from prompts we wrote.

On work covered by a confidentiality agreement, that agreement binds the tools too. You can ask which work went into which tool; we will tell you.

Copyright

You own what we deliver.

We deliver nothing AI-generated in its raw state; every piece passes through our hands, gets edited, and is pulled into the brand's language.

At delivery we state in writing which part of a piece was produced how. If you want it, it goes into the contract too.

Where we don't use it is part of this statement too.

In strategy.: A model can't decide what a brand should say. That conversation happens at the table, with you.
In the final word.: No line and no frame goes live without passing approval.
In thinking on your behalf.: We use it to move faster, not to stand in for the work.

If you'd rather we didn't, we don't.

You can ask for AI to be out of every stage of your project. We talk it through in the first conversation.

The timeline changes. The budget changes. Our answer doesn't: yes.

We write down what didn't work, and what it costs.

Where we stopped

We tried it, it didn't work, we stopped.

We tried it for copywriting. The sentences were correct; the voice wasn't ours. We stopped.

Saying that something didn't work is part of this statement too.

The cost

It doesn't come free.

Every prompt spends energy. A studio that argues for light websites would be inconsistent to ignore that.

So we don't treat AI as a playground. We ask knowing what we are asking; we don't generate ten variants and pick one.

Fewer prompts, better prompts.

Sustainable Web

What we do, and what we don't.

We do

Refactoring, tests, debugging, security scanning
Accessibility and performance audits
Comparing alternatives quickly
Raw visual and motion generation

We don't

Have a model build the whole site
Ask a model for brand strategy
Publish a single line that hasn't been approved
Have text generated as if it were your voice

Three questions, three short answers.

Is my data used to train models?

It depends on the tool, and we write down both answers. Claude, which we use for code and research, and Raylight, which we use for interface videos, do not open data to model training; Higgsfield, which we use for visuals and video, does on its standard plan. That is why your brand files, your code and your unreleased work are never uploaded to Higgsfield. On work covered by a confidentiality agreement, that agreement binds the tools too.

Who owns the copyright on AI-generated visuals?

You do. You own everything we deliver; we hand nothing over in its raw state, and at delivery we state in writing which part of a piece was produced how.

Can I ask you not to use it on my project at all?

Yes. You can ask for AI to be out of every stage of your project; we talk it through in the first conversation. The timeline and the budget change, our answer does not.

We make the decisions, we do the work.

Is there something you find missing on this page? Write to us and we'll add it.

Work With Us Sustainable Web The Maneko Way

This statement was last updated on September 2, 2026. Version 1.1.

As it changes we write it down here; we keep the older versions too.

Change log

1.1
- Added the tools we use, as a data-policy table.
- Made the data section honest tool by tool.
- Added the data and copyright section.
- Added evidence in numbers, from our maintenance records.
- Added the sections on your decision, where we stopped, and the energy cost.
- Made the boundaries section the loudest part of the page.
- Added the do / don't summary and frequently asked questions.
- Added a one-line answer under the heading.
- Added the version stamp and this change log.
1.0
- First publication.