chatgpt-compose

name	chatgpt-compose
description	Drop-in alternative to nano-banana/FAL compose — drives the ChatGPT web UI under your Plus/Pro subscription to do multi-ref image composition, single-image edits, and brand bakes. Same CLI shape as nano-compose, no API cost, no FAL credits. Off-label use of browser-automate (ChatGPT has aggressive bot detection — see fragility notes).
metadata	{"clawdbot":{"emoji":"🖼️","requires":["bun","browser-automate skill","browser-automate daemon"],"credentials":"none at the skill level — sign in to chatgpt.com in your daily Chrome (or the daemon-seeded profile) and the daemon inherits the session"}}

chatgpt-compose

Replaces nano-compose for callers that want to use a paid ChatGPT subscription instead of FAL/nano-banana credits. Same flag shape, different backend. Pick per-shot:

# FAL credits path
bun <home>/clawd/skills/nano-banana-compose/bin/nano-compose.ts \
  --output OUT.png --prompt "..." REF1.png REF2.png

# subscription path (ChatGPT web UI, no credits)
bun <home>/clawd/skills/chatgpt-compose/bin/chatgpt-compose.ts \
  --output OUT.png --prompt "..." REF1.png REF2.png

ChatGPT caps 20 attachments per message — so this comfortably handles large ensemble keyframes (6+ character refs + a location).

When to use this vs nano-compose

Use case	Recommended backend	Why
Primary character / location generation (0 refs, text→image) — content-safe prompts	chatgpt ✅	Replaces flux-pro/imagen4 for these; subscription-paid, no per-image cost.
Primary gen — prompts with content-policy edges (cannabis, weapons, etc.)	flux-pro/imagen4 (NOT chatgpt) ❌	DALL-E silently substitutes the prompt with a "safe" alternative instead of refusing. The substituted image gets saved to disk under your character's filename, silently corrupting the locked ref. Use a backend that refuses-on-policy rather than substitutes.
Single-ref restage / pose change (1 ref)	chatgpt ✅	Character identity preserved, scene composed cleanly, ~70s end-to-end
Brand bake (logo placement, merch mockup, ≤2 refs)	chatgpt ✅	DALL-E reproduces logos via image ref reasonably well; cheaper than burning FAL credits
QC auto-correction (single targeted fix)	chatgpt ✅	Cheap, fast, avoids burning FAL credits on small changes
2-character keyframe (2 char + 1 loc refs)	either backend works ✅	Identity, accessories, coat colour all preserved by both. nano edges chatgpt on background plate fidelity; chatgpt frames closer character-portraits.
Ensemble keyframe (5–7 refs)	nano-banana ✅	Side-by-side tests: nano kept more characters with stronger location-plate fidelity. chatgpt drifted more on abstract characters and rendered generic backgrounds. Both drifted on the most abstract refs.
Any text rendering on the image	neither — overlay via ffmpeg post	Both DALL-E and nano-banana garble text

The species-conflation result for ensemble shots matches nano-compose's own SKILL.md known failure mode — DALL-E exhibits it more aggressively. nano-banana/edit's image_urls array conditioning gives stronger per-ref identity weight than DALL-E's text-context-with-attachments approach.

One-time setup

1. Install browser-automate (if not already)

cd <home>/clawd/skills/browser-automate && pnpm install && pnpm run build

2. Sign in to ChatGPT in your daily Chrome

Open https://chatgpt.com/ in your normal Chrome, sign in, complete any 2FA, dismiss the welcome modal. The browser-automate daemon seeds from this profile on first start.

3. (Recommended) Create a Project for chat hygiene

In ChatGPT's left sidebar → + New project → name it e.g. compose-bakes. Open the project, copy the URL — it'll look like https://chatgpt.com/g/g-p-XXXXXXXX/project. Set:

export CHATGPT_START_PATH="/g/g-p-XXXXXXXX/project"

(Or pass --start-path per call.) Every compose invocation will land on the project, so all generated chats stay scoped there instead of cluttering your main sidebar.

4. Start the daemon

<home>/clawd/skills/browser-automate/bin/browser-automate daemon start

This launches a debuggable Chrome that inherits your daily profile (cookies, session). Stays running across calls — leave it up. Stop with daemon stop.

Usage

Zero-ref primary generation (text-to-image, no references)

For primary character/location refs from a bible description. No positional args.

bun <home>/clawd/skills/chatgpt-compose/bin/chatgpt-compose.ts \
  --output ./locations/rooftop-night-locked.png \
  --aspect-ratio 16:9 \
  --prompt "Rooftop balcony, late night. Warm string lights, low velvet seating, distant city skyline below in muted amber haze. Painterly children's-book illustration style. No characters, no text."

Single brand bake (1 ref, logo placement)

bun <home>/clawd/skills/chatgpt-compose/bin/chatgpt-compose.ts \
  --output ./char-with-logo.png \
  --prompt "Emboss the crown logo from the reference onto the medallion. Keep EVERYTHING else identical." \
  ./char-presigil.png ./brand-logo.png

Ensemble scene keyframe (7 refs)

bun <home>/clawd/skills/chatgpt-compose/bin/chatgpt-compose.ts \
  --output ./p1-s03-keyframe.jpg \
  --prompt "Compose all six characters together at the location from the last reference. Close three-quarter framing on the foreground pair, others in soft focus." \
  --aspect-ratio 16:9 \
  ./char-1.png ./char-2.png ./char-3.png ./char-4.png ./char-5.png ./char-6.png ./loc.png

Targeted QC correction

bun <home>/clawd/skills/chatgpt-compose/bin/chatgpt-compose.ts \
  --output ./fixed.png \
  --prompt "Add a small gold pendant to the character's neck. Don't change anything else." \
  ./qc-failed-frame.png

Flags

Flag	Purpose
--output <path>	Where to save the generated image (required)
--prompt "<text>"	Edit / composition instruction (required)
--aspect-ratio <ratio>	e.g. 16:9, 1:1, 9:16 — appended to prompt as natural language
--start-path <path>	URL fragment to land on, e.g. /g/g-p-XXX/project. Defaults to $CHATGPT_START_PATH or /.
--timeout-ms <n>	Image generation timeout (default 300000 = 5 min)
IMG1 IMG2 …	Reference image paths (positional, 0–20). Zero refs = text-to-image generation.

Output (stdout JSON)

{
  "ok": true,
  "provider": "chatgpt-browser",
  "output": "./out.png",
  "bytes": 1820144,
  "imageUrl": "https://files.oaiusercontent.com/...",
  "chatUrl": "https://chatgpt.com/c/...",
  "durationMs": 87421
}

The chatUrl lets you reopen the generation in your browser to inspect, regenerate, or iterate manually.

Prompt discipline

DALL-E benefits from the same rules as nano-banana:

1. Reference inputs by position — "the character from the first reference" / "the location from the last reference"
2. Pin the unchanged parts explicitly — "Keep EVERYTHING else identical — same {features}. Only {the edit}." prevents re-imagining
3. For brand bakes, attach the logo PNG as a ref — reproduces logos faithfully from image input, not text
4. For BG continuity in close-ups — explicit framing direction ("close three-quarter portrait, soft-focus secondary character at frame edge")

Fragility — read before debugging

This skill goes off-label on browser-automate's own warning: ChatGPT sits behind Cloudflare with active fingerprinting. Mitigations baked in:

• CDP-attach via daemon, not Playwright-launched profile — uses your real Chrome with real fingerprint and real session. Same pattern as spotify-creators.
• Composer typed via keyboard.type, not .fill() — ChatGPT's input is a ProseMirror contenteditable; .fill() no-ops on those.
• Network capture for the image URL — DOM scraping the assistant <img> is fragile because the turn re-renders during streaming. The recipe listens on page.on('response') for files.oaiusercontent.com / oaidalleapi*.blob.core.windows.net URLs.
• Refusal detection — if the chat surfaces "content policy" / "can't help with that" text, the recipe fails fast instead of waiting 5 min.

When it breaks

• Cloudflare challenge → browser-automate run chatgpt --headful --args-json '...', solve once manually, retry. Session persists in daemon Chrome.
• no <input type=file> found → ChatGPT moved the paperclip menu. Inspect the page; update SELECTORS.fileInput in <home>/clawd/skills/browser-automate/src/recipes/chatgpt.ts.
• composer not found → same — update SELECTORS.composer.
• send button never enabled → upload silently failed (oversize file, weird format, rate limit). Check error screenshot at /tmp/browser-automate/chatgpt/<ts>/error-*.png.
• no generated image detected within Xms → either the prompt was refused without a content-policy banner, or DALL-E URL pattern moved. Open the chatUrl from a successful prior run to inspect current image hosting domain, then update isGeneratedImageUrl().
• DALL-E quality regression vs nano-banana → expected at higher ref counts. Fall back to nano-compose for that shot.

Account safety

• One-account, one-session use only. Don't parallelise across the same login.
• ChatGPT enforces TWO rate limits, both relevant:
  • Image generation: Plus ~50/3hr, Pro higher. Trips silently — the assistant returns text instead of an image.
  • Conversation creation: creating ~6+ new chats inside ~10 min triggers a "Too many requests / temporarily limited access" modal that blocks all conversation actions. The recipe detects this up-front and fails fast.
• If you trip the conversation rate limit, wait 5–10 min. The session itself is fine; only new-chat creation is throttled.
• For batch workflows (e.g. generating a full character roster), strongly prefer one chat with sequential prompts over N new chats. Reduces both rate-limit risk and sidebar pollution. (The recipe currently creates a new chat per call — TODO: add --continue-chat <id> mode.)

Related

• browser-automate — engine + login + daemon
• nano-banana-compose — original nano-compose, canonical compose-discipline rules
• spotify-creators — same CDP-attach pattern reference