ChatGPT Images browser playbook
Automate ChatGPT Images through a signed-in box browser session. No API keys. No cookie scraping. No private endpoint scripts. No Shell-driven CDP, Playwright, Puppeteer, or cookie DB reads.
This is a session-only playbook: drive the live Images UI the owner already uses. It is product-agnostic — attach whatever identity or style reference images the current job requires.
Session
- Open
https://chatgpt.comin the box browser on the account the owner already uses for Images. - Confirm Images / image gen is available (Plus or whatever plan they use). If login or 2FA blocks, hand the box to the owner. Do not lift cookies.
- Prefer one long signed-in session. Reuse the tab. Do not open parallel browsers that fight the same login.
How to drive the UI
- Delegate clicks and typing to a browser or desktop subagent. Do not drive Chrome from Shell.
- Scope each dispatch tightly: one prompt, attach named refs, wait for the still, download or save it, stop.
- For a batch, loop many narrow dispatches (one still each). Do not ask one subagent to "generate the whole pack."
- Always attach the locked identity / style reference images the job requires before Generate. Do not invent product cast locks or character sheets in this skill — those belong to the project brief.
- Use the image-generation controls available in the current session. Do not assume a specific model name or picker exists. Reject GenerateImage and other local generators for this lane.
- No sleep-poll loops. Do not chain
sleep 15/sleep 20/sleep 30waiting on the UI. Wait on the page result in the browser subagent, or stop and report. Retry once on rate limit, then pause.
Batch recipe
For each item in the current stage list (from the project PLAN or a companion consistency skill):
- Build one clear English prompt (subject, look, camera, what must stay exact).
- Dispatch: open Images, attach refs, paste prompt, generate, wait until the image is ready.
- Save the file into the project folder with a stable name.
- Quality-gate against the project brief. Redo once if it fails. Escalate to the owner after two fails.
- Mark the item done and move to the next. Report progress; offer a question widget to continue or pause.
Hard stops
- Login wall, captcha, phone verify, or payment wall → hand the box to the owner.
- Rate limit or "try again later" → wait and retry once, then pause the batch and tell the owner.
- UI changed so the Images controls cannot be found → stop and report. Do not invent a cookie or API workaround.
- Any urge to scrape session cookies or call undocumented ChatGPT APIs → refuse.
Anti patterns
- GenerateImage "just this once."
- Parallel browser sessions on the same ChatGPT login.
- Sleep-poll shell loops instead of waiting on the page.
- One sprawling "do the whole pack" computerUse task.
- Baking product-specific cast boards or character locks into this playbook.