Generate images from Claude Code

你說

Make a clean minimalist poster of a small robot assistant holding a paint palette — soft studio light, warm beige background.

AGENT 回覆

Image generation. Three options from the catalog, priced per run:

• GPT Image 2$0.15 per rungeneration and precise edit instructions
• Seedream 5 LTE$0.10 per runfast drafts
• Banana 2$0.20 per rungeneration and reference-based edits

你選定: GPT Image 2

• Calls the image service with the GPT Image 2 variant — the run is accepted as a job
• Waits on the job; this one finished in about 75 seconds
• Downloads the image into the project's outputs/ folder and hands you the file

結果

1 run · GPT Image 2 · $0.15 per run · charged from the wallet after completion

That is the entire loop, unedited, from a real session: one sentence in, a file out, one line on the bill. Claude Code itself does not ship an image model — the work goes through Actionway's image generation capability, and the Skill installed with the CLI is what teaches the agent to plan, price, run, and deliver it.

What happened in between

Before answering, the agent checked the capability catalog it keeps locally, so the first reply already carried the plan: which capability, which models, what each costs per run. It presents those options neutrally — no default, no recommendation; say "you pick" and it takes the best-quality tier. Picking one is the approval: the call runs at the shown price, and Actionway checks the wallet on its side. Only runs the platform flags for extra confirmation pause for another question first.

Image generation runs as a job, so the agent waits on it rather than polling by hand, then downloads the finished asset into an outputs/ folder in your project with a descriptive filename. You get a local file, not a link. If the wallet cannot cover a run, the agent stops before anything is charged and gives you a top-up link with the amount prefilled.

Pick a model, pay per run

Three families cover most image work, at fixed catalog prices: GPT Image 2 ($0.15 per run) for generation and precise edits, Seedream (from $0.10 per run) for fast drafts and final renders, and the Banana models ($0.20 per run for Banana 2, $0.30 per run for Banana Pro) for reference-based edits. The prices are rendered from the same registry the models library shows, so they match what your agent quotes.

Editing works like generation: hand the agent an existing image plus the change you want, and it routes the job to a variant that accepts reference input. Every attempt is one catalog-priced run, and there is no subscription — a balance you top up once covers every capability in the catalog.

Writing a brief the model can follow

The prompt in the case is short on purpose: one subject, one style, one light, one background. Image models do best with that shape. Say what the picture is of before you say how it should look, name the surface it will live on if that matters (a square social tile, a wide banner, a portrait poster), and keep any text you want in the image to a few words in quotation marks — long sentences come back misspelled. Leave out what you do not care about; every extra adjective is a constraint the model may satisfy at the expense of the one you did care about.

Iterate on the draft tier. A Seedream 5 LTE run costs $0.10 per run, so trying five phrasings of the same brief is cheaper than one premium run, and the agent keeps every file, so nothing is lost when you go back to a version from three attempts ago. When a draft is close, send the same sentence to GPT Image 2 for the final; if it is close but one detail is wrong, hand the draft back as a reference and describe only the change.

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 reports your wallet balance and, if it is empty, links you straight to top-up. Prefer a plugin or a Remote MCP server instead of a CLI? Both routes end at the same catalog and the same wallet — the get started page lists the steps per client side by side.