AI image prompts, and what the images carry

Prompt guides for the generators people use now, and what each one writes into the finished image.

frame 11Frame 11: clay-covered hands pulling up the wall of a bowl on a potter's wheel, seen from above
▸ 11NF 01
frame 11AFrame 11A: hands steadying a wide clay bowl on the wheel, a water bucket and sponge behind
▸ 11ANF 01
frame 12Frame 12, the kept frame: a potter in a denim apron shapes a bowl on the wheel beside a window
▸ 12NF 01 · KEPT
frame 12AFrame 12A: a potter shapes a small bowl on the wheel in bright window light
▸ 12ANF 01
Prompt · frames 11 to 12A
Overhead photograph of a ceramicist's hands shaping a wet clay bowl on a pottery wheel, 35mm film, soft window light from the left, clay dust in the air, shallow depth of field. Shot as a single 35mm frame, natural key light from one side, fine film grain, true colour, no text, no logos, no watermarks, no borders.

Run on the Gemini API, model gemini-3.1-flash-image (Nano Banana 2) · 2026-11-02 · 3:2 · no seed set

What frame 12 carries

SignalYesNo
Content credentialmarkednot marked
Invisible watermarkmarkednot marked
Metadata fieldsnot markedmarked
Prompt saved in filenot markedmarked

Checked on 2026-11-02 with c2pa-rs, on the file exactly as the API returned it. The credential is signed by Google and records "Applied imperceptible SynthID watermark"; we cannot run Google's SynthID detector ourselves. No EXIF, XMP or IPTC block sits outside the credential, and the prompt is not in the file. The WebP copies on this page carry none of it.

Pick your generator

Prompt guide and file record for each one.

What your AI image carries

The four records a file can hold.

Pick your generator

Each one has a prompt guide and a page on what its finished files carry.

The four things a generated file can carry

diagram · the layers of a fileDiagram of the layers of an image file: pixels, invisible watermark, metadata fields, content credential and saved prompt
▸ 20NF 02
Frame 20. The layers of one file, from the pixels out.

  • Content credential

    Can show:
    which tool signed the file, and the edits it recorded.
    Cannot show:
    anything once the credential is removed from the file.

  • Invisible watermark

    Can show:
    that a vendor's own detector finds its mark in the pixels.
    Cannot show:
    anything to a reader who has no access to that detector.

  • Metadata fields

    Can show:
    what the software wrote into the file, such as a source type.
    Cannot show:
    much on their own, since any editor can change them.

  • Prompt saved in file

    Can show:
    the words and settings a front end recorded.
    Cannot show:
    that this prompt alone made this exact image.

A missing label does not mean an image is human-made.

What your AI image carries

Prompt craft

All prompt guides

Which generators label their images

An excerpt from the full comparison, with its method and sources.

Excerpt: the generators whose files we have read ourselves, as of 2026-11-02.
GeneratorContent credentialInvisible watermarkVisible markPrompt saved in file
OpenAI image APIYes, signed by OpenAIYes per OpenAI's docs; the credential records itNoNo
Gemini API (Nano Banana 2)Yes, signed by GoogleYes, SynthID per the credential and Google's docsNoNo
feature · an image on its way to a shopA generated print mocked up on a product page, next to its declaration form
▸ 31NF 05

Selling and labelling

For images that leave your machine: stock sites, shops and feeds.

Newest guides

From 2023 to 2024 this domain hosted the NeutronField prompt marketplace, which closed in 2024. This site is a new publication and is not affiliated with it.

What happened to prompt marketplaces →

How the guides are written

Every prompt shown here was run by us, with the model and date beside it, and every claim about what a generator writes into a file comes from a file we generated and read ourselves, checked against the vendor's own documentation.

About NeutronField