Claude Code can't make your listing images. Codex can. ⤷ And Claude can call it.
By Jaspinall · November 9, 2026 · Curated by George's Blog
Claude Code can't make your listing images.
Codex can.
⤷ And Claude can call it.
Last year: pick one AI tool & live inside it.
Now: run two agents & let each do the job it's built for.
Here's the setup I'm running for Amazon image stacks:
Claude runs the skill → writes the brief & prompts → hands generation to Codex CLI → Codex makes the images → Claude pulls them back and checks them.
One command does the hand-off: codex exec.
Claude calls it like any other terminal tool.
⛔️ No API key.
⛔️ No new app.
Skip this if you make 3 images a month.
Use it if creative is a weekly line item.
One setting to check first.
If your Codex config runs full access with approvals off, anything Claude hands it can touch your whole disk.
⤷ Lock every job to its run folder (-s workspace-write).
My bet? The operator who wins this year isn't the one who picks the best model. It's the one who wires two together.
Why it works:
① 𝗜𝗺𝗮𝗴𝗲𝘀 𝗼𝗻 𝘆𝗼𝘂𝗿 𝗽𝗹𝗮𝗻. Codex image generation runs on the ChatGPT subscription you already pay for. I still have 14 graphic skills paying per image on an API. They're next.
② 𝗔 𝘀𝗲𝗰𝗼𝗻𝗱 𝗺𝗼𝗱𝗲𝗹 𝗴𝗿𝗮𝗱𝗲𝘀. A model grading its own work passes its own weak work. Claude builds the stack, GPT looks at the finals next to your live carousel and flags the downgrade.
③ 𝗣𝗮𝗿𝗮𝗹𝗹𝗲𝗹 𝗿𝗼𝗹𝗹𝗼𝘂𝘁𝘀. Approved stack, a dozen SKUs to roll it to? Claude fires 2-3 Codex jobs at once and QAs each one as it lands.
④ 𝗥𝗲𝘃𝗶𝘀𝗶𝗼𝗻𝘀 𝗸𝗲𝗲𝗽 𝘁𝗵𝗲 𝗹𝗼𝗼𝗸. "Make the headline bigger" resumes the same Codex session. No regenerating from scratch and losing the style you approved.
⑤ 𝗢𝗹𝗱 𝘀𝗸𝗶𝗹𝗹𝘀 𝗰𝗼𝗺𝗲 𝗯𝗮𝗰𝗸. I had 19 skills living only in Codex. Claude can call every one of them without porting a line.
A $20 entry level paid ChatGPT account can do LAPS around raw API usage.
What do you think?