gpt-image2 - SKILL.md Agent Skill

name: gpt-image2 description: Generate images with GPT Image 2 through a local OpenAI-compatible Images API script, then deliver them through OpenClaw chat channels such as Telegram or QQBot. Use when the user asks to generate, draw, create, or send an image, poster, photo, illustration, wallpaper, avatar, icon, logo, 文生图, 生成图片, 画图, 海报, 照片, 发我图片, or similar visual output requests.

GPT Image 2

Overview

Use this skill to generate images from chat with gpt-image-2 through a local PowerShell script:

scripts/generate_gpt_image2.ps1

The script calls an OpenAI-compatible /v1/images/generations endpoint, saves a PNG locally, and optionally sends it through Telegram or emits a QQBot media payload.

Do not use or mention nonexistent image_gen, openai-image-gen, or nano-banana-pro tools. Do not send gpt-image-2 image requests to a chat/completions endpoint.

Resolve Script Path

If the current directory is not this skill folder, resolve the script path from OpenClaw:

$skillFile = (openclaw skills info gpt-image2 | Select-String 'Path:' | ForEach-Object { ($_ -split 'Path:\s*', 2)[1].Trim() } | Select-Object -First 1)
$script = Join-Path (Split-Path -Parent $skillFile) 'scripts/generate_gpt_image2.ps1'

Channel Workflow

Choose the delivery mode from the active channel or from the user's wording.

For Telegram delivery, run:

pwsh -NoProfile -File $script -Prompt "<image prompt>" -SendTelegram -TelegramTarget "<telegram chat id>" -Caption "<short caption>"

If -TelegramTarget is omitted, the script reads OPENCLAW_TELEGRAM_TARGET. If neither exists, report the error instead of pretending the image was sent.

For QQBot image replies, run:

pwsh -NoProfile -File $script -Prompt "<image prompt>" -EmitQQPayload -Caption "<short caption>"

Then make the final response exactly the QQBOT_PAYLOAD: block printed by the script, with no extra text before or after it.

For local generation without sending, run:

pwsh -NoProfile -File $script -Prompt "<image prompt>"

Then return the saved file path.

Provider Configuration

By default the script reads ~/.openclaw/openclaw.json and uses provider id cpamc, which can be changed with -ProviderId or OPENCLAW_IMAGE_PROVIDER.

Use one of these patterns:

pwsh -NoProfile -File $script -Prompt "<prompt>" -ProviderId "cpamc"
pwsh -NoProfile -File $script -Prompt "<prompt>" -ProviderId "openai"
pwsh -NoProfile -File $script -Prompt "<prompt>" -BaseUrl "https://api.openai.com/v1" -ApiKeyEnv "OPENAI_API_KEY"

Never print API keys, bot tokens, or full private config contents.

Prompting

Expand short user requests into a concrete visual prompt before calling the script. Preserve required subjects, locations, text, style, aspect, and language from the user's request.

Use -Size 1024x1024 unless the user asks for another supported size.

Failure Handling

If the script reports a provider, network, timeout, configuration, or delivery error, include the actionable error. Do not produce an empty answer or a verbal promise without a successful script result.