One product photo, three ad creatives: the Claude Code image editing skill

1. The night-street version

YOU ASK

Edit outputs/sneaker.jpg: place this exact white sneaker on a rain-wet city street at night, neon shop signs reflected in the puddles, low camera angle, keep the sneaker itself unchanged, no logos, no text

AGENT REPLIES

Image edit with a reference. Three models in the catalog accept one, priced per run:

• Banana Pro$0.30 per runreference-based edits, the higher Banana tier
• GPT Image 2$0.15 per rungeneration and precise edit instructions
• Banana 2$0.20 per runreference-based edits, the lower Banana tier

You pick: Banana Pro

• Registers the photo as an asset and calls the image service to Banana Pro — it starts as a background job
• Waits on the job; this one finished in 2 minutes 20 seconds
• Downloads the edited image into an outputs/ folder in your project (that is where every generated file lands) next to the original

RESULT

1 run · Banana Pro · $0.30 per run · charged from the wallet after completion

That is a real Claude Code session, unedited. The shoe is the shoe — same shape, same laces, same gum sole — and everything around it is new. "Keep the sneaker itself unchanged" is the phrase doing the work; without it, edit models tend to redesign the product a little while they change the scene.

2. The flat-lay version

3. The outdoor version

Keeping the product the product

Edit models redraw the whole frame, so the thing you must protect is the product itself. Three phrases did that work in every version above: "this exact sneaker" tells the model the reference is the subject, not inspiration; "keep the sneaker itself unchanged" stops it from tidying the laces or slimming the sole; "no logos, no text" stops it from adding lettering, which edit models do to shoes and walls when nothing forbids it. Then describe the scene in one sentence — surface, light, angle — and stop.

Check each result against the reference before you approve a batch. Compare the sole colour, the lace count, the height of the collar, and the stitching line along the toe: those are the details that drift first. A version that drifted is one more run on the cheaper tier, and the agent keeps the reference and the prompt, so the retry is a sentence: "same again, but keep the gum sole".

How the session decides what to run

Claude Code can read an image file and tell you what is in it; producing a changed version is outside what it does. The Skill installed with the CLI closes that gap: before replying, the agent read the capability index the Skill keeps on disk and saw which image models take a reference. That is why the first message already listed three of them with prices — neutrally, no default, no recommendation. Say "you pick" and it takes the best-quality tier. Picking a model is the approval: the run starts at the shown price with no second question, and if the wallet cannot cover it the agent stops before anything is charged and gives you a top-up link with the amount prefilled.

Two families take a reference image, at fixed per-run prices. The Banana models — Banana Pro at $0.30 per run, Banana 2 at $0.20 per run — are built around reference-based edits. GPT Image 2 at $0.15 per run handles generation and edits from precise instructions in one model. For a campaign, run the cheaper tier for the first pass of each idea, then spend on the versions you keep. The video generation services take a reference image too, which is how the same photo becomes a five-second product clip in the same session. Every edit went through Actionway's image generation capability.

Set it up once

npm install --global @actionway/cli@latest
actionway init
mkdir -p "$HOME/.claude/skills"
cp -R "$HOME/.actionway/skill/actionway" "$HOME/.claude/skills/actionway"

You need Node.js 20 or newer and an Actionway account; after sign-in the agent tells you your wallet balance and, if it is empty, links you straight to top-up. Prefer a plugin or a Remote MCP server to the command line? Both end at the same catalog and the same wallet — the get started page lists the steps per client side by side.