A model writes the data. Your browser draws the picture.

What the model actually does

When you describe a sprite, your sentence goes to Claude, made by Anthropic. It does not send back a picture. It sends back a list: an ordered stack of shapes, a table of colours, and frames if the thing moves. Your own browser then draws that list into pixels.

This is the difference between Bitwright and a tool that generates images. The model never touches a pixel. It writes a specification, and a renderer that lives in the page turns it into a sprite, which is why every part is still there to be switched off and every colour is still there to be changed.

What is sent, and what is not

Sent to Anthropic:

• The description you typed.
• The viewing angle and whether you asked for movement.
• Any sprite you asked it to match, or any photograph you uploaded.
• A fixed instruction sheet describing the format, identical for everybody.

Not sent:

• Your email address, your account, or anything identifying you.
• Anything you drew previously, unless you explicitly attach it as a match.

Anthropic processes this under their commercial terms, which means it is not used to train their models. We do not store your description or the result at either end; see the privacy policy for what is kept, which is a row saying what it cost.

Who owns the result

You do, on any account, including commercially and with no credit required. We claim nothing. The terms say this properly, along with the one exception: the single free draw without an account is a demonstration rather than a licence.

Two honest caveats, true of every tool of this kind:

• Copyright in purely machine-generated work is unsettled in most countries, so nobody can promise a sprite is protectable. We can only promise we will not claim it.
• Similar descriptions produce similar sprites. Nothing here is guaranteed unique.

Where it goes wrong

A model draws these, so it will sometimes misunderstand you. Being straight about the failure modes is more useful than claiming there are none.

• It can ignore part of your description. Two-object requests are the usual case: a mug with a flower in it may come back as a very good mug.
• Detail has a floor. At thirty-two pixels across, a face or small text becomes a suggestion. Asking for finer detail than the canvas holds will not produce it.
• Characters are weaker than objects. The format was built for props and furniture. People and creatures work less well and that is a known gap rather than a temporary bug.
• An allowance is spent either way. A sprite you did not like cost exactly the same to produce as one you did, which is why it counts.

What we will not draw

The model will decline some requests on Anthropic's own safety rules, and we add our own in the acceptable use policy. In short: nothing illegal, nothing sexual involving minors, nothing built to harass a real person, and no deliberate copying of somebody else's trademark or character.

This will change

The model and its settings are configuration rather than a promise, and they get changed when a comparison shows something better or cheaper. Every drawing records which model and setting produced it, so a change can be measured rather than guessed at. If a change alters what you get in a way you would notice, this page will say so.