The Headphones Were in the Brief

An image agent will finish your brief faster than you can. It will also finish every gap in that brief, at exactly the same speed.

We watched it happen on Thursday, October 8.

Israel Ayliffe ran one character study several ways at once. By hand in ChatGPT, on GPT Image 2.5 with thinking set to extra high. Through Muse, an AI agent working from a shared brief inside Flora. Through Google Flow on Nano Banana 2.1, for comparison.

The brief gave the model four reference images and a job for each one. The fourth, on screen in the image above: "Fourth attachment: frontal pose, headphones over ears, hands in jeans pockets. Frontal proportions anchor."

Hold that thought.

What does an image agent do with your brief?

Everything it says.

The headphones turned up in almost every frame, whatever the scene. Muse kept going. Israel stopped and handed the problem to Codex: read the prompts and find what is asking for headphones.

Codex found two causes. The prompts requested headphones in almost every outfit. And the instruction to preserve the character told the model that accessories were part of who this person is.

Israel's read on it, live: the AI "wouldn't think of doing it itself. It would just lock on to the brief I've given it."

He told Muse to start over with the headphones taken out. Muse reported them gone from all 25 prompts.

That is the move worth stealing. When the output goes wrong the same way twice, read the instructions before you touch the model.

Hand or agent, which did better?

Each one won a different race.

Muse was quick, and it took liberties that paid off. On one frame it tipped the camera into a Dutch angle and gave the character an attitude the Nano Banana version never found.

By hand, Israel fixed things as they appeared, one image at a time. Slower, with every frame passing through a person with an opinion. The agent works the brief. The hand works the taste. In his words, "if your brief is perfect, AI will get it done perfectly." His verdict at the close: given time, editing as you go gives the best results.

Nobody's first brief is perfect. That is where the hand earns its keep.

Natural language or JSON prompts?

Both work. Israel ran the same illustration-to-photo job both ways, and his read was that natural language gets you there and JSON gets you there too.

His JSON prompts spelled out the plan, a reference lock, the subject, pose, wardrobe, environment, camera, lighting, style, and how to check the result. He was straight about it: some people find that overwhelming.

The brief also said what must not move. Keep the face. Don't beautify the subject. Write it down in words.

What to carry into your own image work

• When output goes wrong the same way twice, read the brief first.
• Write down what must never change, in words.
• Give an agent a stopping point and a budget. Credits run out mid-job, and Israel checked usage in Flora and Muse live while the jobs ran.
• Choose by hand. The agent makes the pile. A person picks.

Frequently asked questions

Is an image agent better than generating by hand?
Faster, yes. On October 8 the agent won on speed and variety, and hand editing gave Israel control of every frame.

Do I need JSON prompts?
No. Israel showed both, and both worked. JSON spells more out, and some people find it overwhelming.

Where is the template Israel used?
In the session guide, shared in our Discord, where updates get posted as he revises it.

See Muse run a real job

On Saturday, October 17, Israel walks through a job Muse ran overnight, start to finish, including the calls only a person can make. It Ran All Night: Muse with Israel Ayliffe, 10:00 to 11:30 AM ET. $39, and included with Creator membership.

Everything we run: https://luma.com/TheHouseofCuriosity

The conversation carries on in our Discord, free: https://discord.gg/GYdEYWzpuZ