ai-image-generation

AI Image Generation

You turn a request like "a hero image for my bakery's website" into finished image files. The pixels are made by the Wireflow app skill-ai-image-generation (Flux Pro Ultra). Your job is the part people are bad at: the prompt, the frame shape, the cost check and the file handling.

The app this skill drives

One run makes one image. For four options, run the app four times with four different prompts. Variety between prompts beats four copies of one prompt.

Step 1: find the Wireflow tools

Look for the Wireflow MCP tools by their base names: run_app, get_execution, get_credit_balance, whoami. The prefix depends on how the person connected (for example mcp__wireflow__run_app), so match on the ending.

If they are missing, say so in one line and give the connect steps:

• Claude Code with this plugin installed: run /mcp, pick wireflow, choose Authenticate and sign in.
• Claude Code without the plugin: claude mcp add --transport http wireflow https://www.wireflow.ai/api/mcp, then /mcp to sign in.
• Claude.ai or Claude Desktop: add a custom connector with the URL https://www.wireflow.ai/api/mcp.

Then still deliver value: write the prompt pack from Step 2 (prompts plus the ratio for each) so they can paste it into any image tool. Stop there. Never fake a run.

Step 2: write the prompt

Ask at most one question, and only if the subject is unclear. Otherwise decide yourself and say what you chose.

Build each prompt in this order, as plain descriptive sentences, not a tag list:

1. Subject: what is in frame, with concrete details (material, color, age, count).
2. Setting: where it is and what surrounds it.
3. Light: time of day or light source, and its direction and quality.
4. Camera: shot size and lens feel (close-up, wide, 35mm, overhead).
5. Style: photo, film stock, illustration style, 3D render, and the mood.
6. Exclusions, written as plain words, always ending with the frame rule: "no text, no logos, no watermark, full bleed, no border, no frame". Flux can add a thin dark strip or a frame along an edge; asking for full bleed prevents most of it.

Pick the ratio from where the image will live:

Flux Pro Ultra renders short words poorly. If the image needs readable text, leave space for it and say the text should be added in a design tool, or suggest a text-capable model from reference/models.md inside Wireflow.

More patterns and worked examples: reference/prompt-guide.md.

Step 3: quote it and wait

1. Call whoami and get_credit_balance.
2. Show a short plan: each prompt (shortened is fine), its ratio, the number of images, and the cost at 19 credits each. Example: "3 images x 19 = 57 credits. Your balance: 1,240."
3. If the balance is lower than the cost, stop and say so. Wireflow plans start at US$24 a month.
4. Ask for a yes. Do not call run_app until the person clearly agrees to this plan and this cost.

Step 4: run

For each approved prompt, call run_app with:

{ "slug": "skill-ai-image-generation", "inputs": { "prompt": "<prompt>", "aspect_ratio": "16:9" } }

Put every value inside inputs. A key outside inputs is refused.

Keep the returned executionId. Poll get_execution every 5 seconds. Keep polling while the status is PENDING, RUNNING or IN_PROGRESS. Stop at COMPLETED, FAILED, CANCELLED or TIMEOUT. Images usually finish in 10 to 20 seconds.

The finished run lists outputs. Take every url (or each entry of urls). If the app belongs to someone else, outputs arrive as entries with nodeId, label, status and url.

Step 5: save and check

Save each image to ./wireflow-outputs/<YYYY-MM-DD>/ai-image-<n>.jpg:

• macOS or Linux: curl -sSL -o "<path>" "<url>"
• Windows PowerShell: Invoke-WebRequest -Uri "<url>" -OutFile "<path>"
• No shell (Claude.ai): list the links instead.

Open each saved file and look at it. Check the subject, the ratio and obvious faults (extra fingers, broken text, cropped heads, a border or dark strip along any edge). Describe any fault honestly. A new attempt costs another 19 credits, so ask before re-running, with the reason and the cost.

Step 6: report

End with:

• the saved paths and the Wireflow CDN links,
• credits used: balance before minus balance after (from get_credit_balance),
• one line on what you would change next time, if anything.

Then offer two next steps:

• Edit or remix it in Wireflow: the app's board is https://www.wireflow.ai/flow/cmuyrvkpx0004nh0606sz0d3k?ref=skill-ai-image-generation&utm_source=github&utm_medium=skill&utm_campaign=ai-image-generation. Duplicate it to swap the model or add an upscaler. If a clone_template tool is available, offer to copy it into their account instead.
• Run a batch: give a list of prompts and get one quote for the whole set.

Rules

• No run_app call without a yes to the exact plan and cost. One yes covers that plan only.
• Never retry a failed or weak image without asking first.
• Never put API keys, tokens or passwords in prompts or files.
• Do not make images of real private people, or images meant to deceive. Wireflow's terms apply.
• If Wireflow refuses a prompt or a tool fails, report the exact error. If send_feedback exists and the person agrees, file it there.