chatgpt-image-gen

Generate an image from a text description with ChatGPT via ego-browser, download it, and save it into the project. The whole happy path is one bundled script; this skill mainly covers how to call it, how to handle login, and how to finish up cleanly.

Prerequisites

• ego lite installed and ego-browser on PATH (if ego-browser: command not found, read the ego-browser skill's references/install.md first).
• The user is logged into ChatGPT inside ego lite. If not, the script exits with code 42 — follow the login handoff below.

Runtime notes (important)

The installed ego lite runtime (0.4.4.x) preloads the helper API — useOrCreateTaskSpace, openOrReuseTab, fillInput, click, js, wait, pageInfo, cliLog, completeTaskSpace — not the page / browser / taskSpaces facades some ego-browser docs describe. Also: wait(...) and timeout are in seconds, cliLog(...) is the output channel, and the embedded node runtime does not inherit shell environment variables (pass data via files, like the script does, or interpolate into the heredoc). The bundled script already follows these rules — run it as-is instead of rewriting it against another API.

Workflow

1. Decide the output path. Save into the current working directory by default (the user said "保存到项目目录"). Derive a concise filename from the image description in the user's language, e.g. 西安旅游攻略图.png for "帮我生成一张西安旅游攻略图". If the user gave a name or directory, use that. Default extension .png (ChatGPT returns PNG).

1. Run the generator (takes 1–5 minutes; image generation is slow — set the Bash timeout to 300s):

bash bash <this-skill-dir>/scripts/gen-image.sh "<image description>" "<absolute output path>"

It reuses the task space chatgpt image generation, opens a fresh ChatGPT chat, sends the description, polls until the image appears, downloads it with the page's session cookies, and writes the file. Final line of output is a JSON status.

1. Handle the exit code:

• 0 with {"status":"ok", ...} — continue to step 4.
• 42 (login_required) — hand the browser to the user so they can log in:

bash ego-browser nodejs <<'EOF' const t = await useOrCreateTaskSpace('chatgpt image generation') const h = await handOffTaskSpace(t.id) cliLog(JSON.stringify(h)) EOF Check h.done, then tell the user to log into ChatGPT in the ego lite window and say when done. Only after they confirm, take control back with a new ego-browser nodejs heredoc using takeOverTaskSpace('chatgpt image generation'), then rerun the generator.

• 1 with image_timeout — generation may still be running; open the conversation URL from the output, check visually with captureScreenshot(), and if the image has since appeared, fetch it manually (the poll + download snippet in the script is the reference). Otherwise report the failure.

1. Verify the saved image — check it exists, has non-trivial size, and visually inspect it (e.g. read the image file) to confirm the content matches the description.

1. Close the task space only after verification passed. Use the bundled idempotent cleanup script instead of composing a CLI command or calling completeTaskSpace directly:

bash bash <this-skill-dir>/scripts/close-task-space.sh

{"status":"closed", ...} means it closed the task space. {"status":"already_closed", ...} is also success: a previous cleanup already removed it. Do not run ego-browser --help as a cleanup workaround.

1. Report the saved file path and the ChatGPT conversation URL (from the script output) to the user.

Notes

• Each run sends exactly one image request and waits for it. If the user wants several images, rerun the script per description rather than batching prompts in one chat — separate runs keep filenames and failure handling clean.
• The script opens https://chatgpt.com/ fresh each run, so every image starts a new chat. The ChatGPT conversation remains in the user's history.
• Do not retry blindly on transient failure: read the JSON status first, it distinguishes login, timeout, and download problems.